MapReduce Book Final
MapReduce Book Final
MapReduce Book Final
This is the pre-production manuscript of a book in the Morgan & Claypool Synthesis
Lectures on Human Language Technologies. Anticipated publication date is mid-2010.
ii
Contents
Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ii
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
2 MapReduce Basics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
CHAPTER 1
Introduction
MapReduce [45] is a programming model for expressing distributed computations on
massive amounts of data and an execution framework for large-scale data processing
on clusters of commodity servers. It was originally developed by Google and built on
well-known principles in parallel and distributed processing dating back several decades.
MapReduce has since enjoyed widespread adoption via an open-source implementation
called Hadoop, whose development was led by Yahoo (now an Apache project). Today,
a vibrant software ecosystem has sprung up around Hadoop, with significant activity
in both industry and academia.
This book is about scalable approaches to processing large amounts of text with
MapReduce. Given this focus, it makes sense to start with the most basic question:
Why? There are many answers to this question, but we focus on two. First, “big data”
is a fact of the world, and therefore an issue that real-world systems must grapple with.
Second, across a wide range of text processing applications, more data translates into
more effective algorithms, and thus it makes sense to take advantage of the plentiful
amounts of data that surround us.
Modern information societies are defined by vast repositories of data, both public
and private. Therefore, any practical application must be able to scale up to datasets
of interest. For many, this means scaling up to the web, or at least a non-trivial frac-
tion thereof. Any organization built around gathering, analyzing, monitoring, filtering,
searching, or organizing web content must tackle large-data problems: “web-scale” pro-
cessing is practically synonymous with data-intensive processing. This observation ap-
plies not only to well-established internet companies, but also countless startups and
niche players as well. Just think, how many companies do you know that start their
pitch with “we’re going to harvest information on the web and. . . ”?
Another strong area of growth is the analysis of user behavior data. Any operator
of a moderately successful website can record user activity and in a matter of weeks (or
sooner) be drowning in a torrent of log data. In fact, logging user behavior generates
so much data that many organizations simply can’t cope with the volume, and either
turn the functionality off or throw away data after some time. This represents lost
opportunities, as there is a broadly-held belief that great value lies in insights derived
from mining such data. Knowing what users look at, what they click on, how much
time they spend on a web page, etc. leads to better business decisions and competitive
advantages. Broadly, this is known as business intelligence, which encompasses a wide
range of technologies including data warehousing, data mining, and analytics.
2 CHAPTER 1. INTRODUCTION
How much data are we talking about? A few examples: Google grew from pro-
cessing 100 TB of data a day with MapReduce in 2004 [45] to processing 20 PB a day
with MapReduce in 2008 [46]. In April 2009, a blog post1 was written about eBay’s
two enormous data warehouses: one with 2 petabytes of user data, and the other with
6.5 petabytes of user data spanning 170 trillion records and growing by 150 billion new
records per day. Shortly thereafter, Facebook revealed2 similarly impressive numbers,
boasting of 2.5 petabytes of user data, growing at about 15 terabytes per day. Petabyte
datasets are rapidly becoming the norm, and the trends are clear: our ability to store
data is fast overwhelming our ability to process what we store. More distressing, in-
creases in capacity are outpacing improvements in bandwidth such that our ability to
even read back what we store is deteriorating [91]. Disk capacities have grown from tens
of megabytes in the mid-1980s to about a couple of terabytes today (several orders of
magnitude). On the other hand, latency and bandwidth have improved relatively little:
in the case of latency, perhaps 2× improvement during the last quarter century, and
in the case of bandwidth, perhaps 50×. Given the tendency for individuals and organi-
zations to continuously fill up whatever capacity is available, large-data problems are
growing increasingly severe.
Moving beyond the commercial sphere, many have recognized the importance of
data management in many scientific disciplines, where petabyte-scale datasets are also
becoming increasingly common [21]. For example:
further study. Given the fundamental tenant in modern genetics that genotypes
explain phenotypes, the impact of this technology is nothing less than transfor-
mative [103]. The European Bioinformatics Institute (EBI), which hosts a central
repository of sequence data called EMBL-bank, has increased storage capacity
from 2.5 petabytes in 2008 to 5 petabytes in 2009 [142]. Scientists are predicting
that, in the not-so-distant future, sequencing an individual’s genome will be no
more complex than getting a blood test today—ushering a new era of personalized
medicine, where interventions can be specifically targeted for an individual.
of this phrase and tally up the words that appear to the left. This simple strategy works
surprisingly well, and has become known as the redundancy-based approach to question
answering. It capitalizes on the insight that in a very large text collection (i.e., the
web), answers to commonly-asked questions will be stated in obvious ways, such that
pattern-matching techniques suffice to extract answers accurately.
Yet another example concerns smoothing in web-scale language models [25]. A
language model is a probability distribution that characterizes the likelihood of observ-
ing a particular sequence of words, estimated from a large corpus of texts. They are
useful in a variety of applications, such as speech recognition (to determine what the
speaker is more likely to have said) and machine translation (to determine which of
possible translations is the most fluent, as we will discuss in Section 6.4). Since there
are infinitely many possible strings, and probabilities must be assigned to all of them,
language modeling is a more challenging task than simply keeping track of which strings
were seen how many times: some number of likely strings will never be encountered,
even with lots and lots of training data! Most modern language models make the Markov
assumption: in a n-gram language model, the conditional probability of a word is given
by the n − 1 previous words. Thus, by the chain rule, the probability of a sequence of
words can be decomposed into the product of n-gram probabilities. Nevertheless, an
enormous number of parameters must still be estimated from a training corpus: poten-
tially V n parameters, where V is the number of words in the vocabulary. Even if we
treat every word on the web as the training corpus from which to estimate the n-gram
probabilities, most n-grams—in any language, even English—will never have been seen.
To cope with this sparseness, researchers have developed a number of smoothing tech-
niques [35, 102, 79], which all share the basic idea of moving probability mass from
observed to unseen events in a principled manner. Smoothing approaches vary in ef-
fectiveness, both in terms of intrinsic and application-specific metrics. In 2007, Brants
et al. [25] described language models trained on up to two trillion words.4 Their ex-
periments compared a state-of-the-art approach known as Kneser-Ney smoothing [35]
with another technique the authors affectionately referred to as “stupid backoff”.5 Not
surprisingly, stupid backoff didn’t work as well as Kneser-Ney smoothing on smaller
corpora. However, it was simpler and could be trained on more data, which ultimately
yielded better language models. That is, a simpler technique on more data beat a more
sophisticated technique on less data.
4 As an aside, it is interesting to observe the evolving definition of large over the years. Banko and Brill’s paper
in 2001 was titled Scaling to Very Very Large Corpora for Natural Language Disambiguation, and dealt with
a corpus containing a billion words.
5 As in, so stupid it couldn’t possibly work.
6 CHAPTER 1. INTRODUCTION
For better or for worse, it is often difficult to untangled MapReduce and large-data
processing from the broader discourse on cloud computing. True, there is substantial
promise in this new paradigm of computing, but unwarranted hype by the media and
popular sources threatens its credibility in the long run. In some ways, cloud computing
6 This title was inspired by a classic article titled The Unreasonable Effectiveness of Mathematics in the Natural
Sciences [155]. This is somewhat ironic in that the original article lauded the beauty and elegance of mathe-
matical models in capturing natural phenomena, which is the exact opposite of the data-driven approach.
7 On Exactitude in Science [23]. A similar exchange appears in Chapter XI of Sylvie and Bruno Concluded by
is simply brilliant marketing. Before clouds, there were grids, and before grids, there
8
were vector supercomputers, each having claimed to be the best thing since sliced bread.
So what exactly is cloud computing? This is one of those questions where ten
experts will give eleven different answers; in fact, countless papers have been written
simply to attempt to define the term (e.g., [9, 31, 149], just to name a few examples).
Here we offer up our own thoughts and attempt to explain how cloud computing relates
to MapReduce and data-intensive processing.
At the most superficial level, everything that used to be called web applications has
been rebranded to become “cloud applications”, which includes what we have previously
called “Web 2.0” sites. In fact, anything running inside a browser that gathers and stores
user-generated content now qualifies as an example of cloud computing. This includes
social-networking services such as Facebook, video-sharing sites such as YouTube, web-
based email services such as Gmail, and applications such as Google Docs. In this
context, the cloud simply refers to the servers that power these sites, and user data is
said to reside “in the cloud”. The accumulation of vast quantities of user data creates
large-data problems, many of which are suitable for MapReduce. To give two concrete
examples: a social-networking site analyzes connections in the enormous globe-spanning
graph of friendships to recommend new connections. An online email service analyzes
messages and user behavior to optimize ad selection and placement. These are all large-
data problems that have been tackled with MapReduce.9
Another important facet of cloud computing is what’s more precisely known as
utility computing [129, 31]. As the name implies, the idea behind utility computing
is to treat computing resource as a metered service, like electricity or natural gas.
The idea harkens back to the days of time-sharing machines, and in truth isn’t very
different from this antiquated form of computing. Under this model, a “cloud user” can
dynamically provision any amount of computing resources from a “cloud provider” on
demand and only pay for what is consumed. In practical terms, the user is paying for
access to virtual machine instances that run a standard operating system such as Linux.
Virtualization technology (e.g., [15]) is used by the cloud provider to allocate available
physical resources and enforce isolation between multiple users that may be sharing the
8 What is the difference between cloud computing and grid computing? Although both tackle the fundamental
problem of how best to bring computational resources to bear on large and difficult problems, they start
with different assumptions. Whereas clouds are assumed to be relatively homogeneous servers that reside in a
datacenter or are distributed across a relatively small number of datacenters controlled by a single organization,
grids are assumed to be a less tightly-coupled federation of heterogeneous resources under the control of distinct
but cooperative organizations. As a result, grid computing tends to deal with tasks that are coarser-grained,
and must deal with the practicalities of a federated environment, e.g., verifying credentials across multiple
administrative domains. Grid computing has adopted a middleware-based approach for tackling many of these
challenges.
9 The first example is Facebook, a well-known user of Hadoop, in exactly the manner as described [68]. The
second is, of course, Google, which uses MapReduce to continuously improve existing algorithms and to devise
new algorithms for ad selection and placement.
8 CHAPTER 1. INTRODUCTION
same hardware. Once one or more virtual machine instances have been provisioned, the
user has full control over the resources and can use them for arbitrary computation.
Virtual machines that are no longer needed are destroyed, thereby freeing up physical
resources that can be redirected to other users. Resource consumption is measured in
some equivalent of machine-hours and users are charged in increments thereof.
Both users and providers benefit in the utility computing model. Users are freed
from upfront capital investments necessary to build datacenters and substantial reoccur-
ring costs in maintaining them. They also gain the important property of elasticity—as
demand for computing resources grow, for example, from an unpredicted spike in cus-
tomers, more resources can be seamlessly allocated from the cloud without an inter-
ruption in service. As demand falls, provisioned resources can be released. Prior to the
advent of utility computing, coping with unexpected spikes in demand was fraught with
challenges: under-provision and run the risk of service interruptions, or over-provision
and tie up precious capital in idle machines that are depreciating.
From the utility provider point of view, this business also makes sense because
large datacenters benefit from economies of scale and can be run more efficiently than
smaller datacenters. In the same way that insurance works by aggregating risk and re-
distributing it, utility providers aggregate the computing demands for a large number
of users. Although demand may fluctuate significantly for each user, overall trends in
aggregate demand should be smooth and predictable, which allows the cloud provider
to adjust capacity over time with less risk of either offering too much (resulting in in-
efficient use of capital) or too little (resulting in unsatisfied customers). In the world of
utility computing, Amazon Web Services currently leads the way and remains the dom-
inant player, but a number of other cloud providers populate a market that is becoming
increasingly crowded. Most systems are based on proprietary infrastructure, but there
is at least one, Eucalyptus [111], that is available open source. Increased competition
will benefit cloud users, but what direct relevance does this have for MapReduce? The
connection is quite simple: processing large amounts of data with MapReduce requires
access to clusters with sufficient capacity. However, not everyone with large-data prob-
lems can afford to purchase and maintain clusters. This is where utility computing
comes in: clusters of sufficient size can be provisioned only when the need arises, and
users pay only as much as is required to solve their problems. This lowers the barrier
to entry for data-intensive processing and makes MapReduce much more accessible.
A generalization of the utility computing concept is “everything as a service”,
which is itself a new take on the age-old idea of outsourcing. A cloud provider offering
customers access to virtual machine instances is said to be offering infrastructure as a
service, or IaaS for short. However, this may be too low level for many users. Enter plat-
form as a service (PaaS), which is a rebranding of what used to be called hosted services
in the “pre-cloud” era. Platform is used generically to refer to any set of well-defined
1.2. BIG IDEAS 9
services on top of which users can build applications, deploy content, etc. This class of
services is best exemplified by Google App Engine, which provides the backend data-
store and API for anyone to build highly-scalable web applications. Google maintains
the infrastructure, freeing the user from having to backup, upgrade, patch, or otherwise
maintain basic services such as the storage layer or the programming environment. At
an even higher level, cloud providers can offer software as a service (SaaS), as exem-
plified by Salesforce, a leader in customer relationship management (CRM) software.
Other examples include outsourcing an entire organization’s email to a third party,
which is commonplace today.
What does this proliferation of services have to do with MapReduce? No doubt
that “everything as a service” is driven by desires for greater business efficiencies, but
scale and elasticity play important roles as well. The cloud allows seamless expansion of
operations without the need for careful planning and supports scales that may otherwise
be difficult or cost-prohibitive for an organization to achieve. Cloud services, just like
MapReduce, represents the search for an appropriate level of abstraction and beneficial
divisions of labor. IaaS is an abstraction over raw physical hardware—an organization
might lack the capital, expertise, or interest in running datacenters, and therefore pays
a cloud provider to do so on its behalf. The argument applies similarly to PaaS and
SaaS. In the same vein, the MapReduce programming model is a powerful abstraction
that separates the what from the how of data-intensive processing.
Scale “out”, not “up”. For data-intensive workloads, a large number of commodity
low-end servers (i.e., the scaling “out” approach) is preferred over a small number of
high-end servers (i.e., the scaling “up” approach). The latter approach of purchasing
symmetric multi-processing (SMP) machines with a large number of processor sockets
(dozens, even hundreds) and a large amount of shared memory (hundreds or even thou-
sands of gigabytes) is not cost effective, since the costs of such machines do not scale
linearly (i.e., a machine with twice as many processors is often significantly more than
twice as expensive). On the other hand, the low-end server market overlaps with the
10 CHAPTER 1. INTRODUCTION
high-volume desktop computing market, which has the effect of keeping prices low due
to competition, interchangeable components, and economies of scale.
Barroso and Hölzle’s recent treatise of what they dubbed “warehouse-scale com-
puters” [18] contains a thoughtful analysis of the two approaches. The Transaction
Processing Council (TPC) is a neutral, non-profit organization whose mission is to
establish objective database benchmarks. Benchmark data submitted to that organiza-
tion are probably the closest one can get to a fair “apples-to-apples” comparison of cost
and performance for specific, well-defined relational processing applications. Based on
TPC-C benchmark results from late 2007, a low-end server platform is about four times
more cost efficient than a high-end shared memory platform from the same vendor. Ex-
cluding storage costs, the price/performance advantage of the low-end server increases
to about a factor of twelve.
What if we take into account the fact that communication between nodes in
a high-end SMP machine is orders of magnitude faster than communication between
nodes in a commodity network-based cluster? Since workloads today are beyond the
capability of any single machine (no matter how powerful), the comparison is more ac-
curately between a smaller cluster of high-end machines and a larger cluster of low-end
machines (network communication is unavoidable in both cases). Barroso and Hölzle
model these two approaches under workloads that demand more or less communication,
and conclude that a cluster of low-end servers approaches the performance of the equiv-
alent cluster of high-end servers—the small performance gap is insufficient to justify the
price premium of the high-end servers. For data-intensive applications, the conclusion
appears to be clear: scaling “out” is superior to scaling “up”, and therefore most existing
implementations of the MapReduce programming model are designed around clusters
of low-end commodity servers.
Capital costs in acquiring servers is, of course, only one component of the total
cost of delivering computing capacity. Operational costs are dominated by the cost of
electricity to power the servers as well as other aspects of datacenter operations that
are functionally related to power: power distribution, cooling, etc. [67, 18]. As a result,
energy efficiency has become a key issue in building warehouse-scale computers for
large-data processing. Therefore, it is important to factor in operational costs when
deploying a scale-out solution based on large numbers of commodity servers.
Datacenter efficiency is typically factored into three separate components that
can be independently measured and optimized [18]. The first component measures how
much of a building’s incoming power is actually delivered to computing equipment, and
correspondingly, how much is lost to the building’s mechanical systems (e.g., cooling,
air handling) and electrical infrastructure (e.g., power distribution inefficiencies). The
second component measures how much of a server’s incoming power is lost to the power
supply, cooling fans, etc. The third component captures how much of the power delivered
1.2. BIG IDEAS 11
to computing components (processor, RAM, disk, etc.) is actually used to perform useful
computations.
Of the three components of datacenter efficiency, the first two are relatively
straightforward to objectively quantify. Adoption of industry best-practices can help
datacenter operators achieve state-of-the-art efficiency. The third component, however,
is much more difficult to measure. One important issue that has been identified is the
non-linearity between load and power draw. That is, a server at 10% utilization may
draw slightly more than half as much power as a server at 100% utilization (which
means that a lightly-loaded server is much less efficient than a heavily-loaded server).
A survey of five thousand Google servers over a six-month period shows that servers
operate most of the time at between 10% and 50% utilization [17], which is an energy-
inefficient operating region. As a result, Barroso and Hölzle have advocated for research
and development in energy-proportional machines, where energy consumption would
be proportional to load, such that an idle processor would (ideally) consume no power,
but yet retain the ability to power up (nearly) instantaneously in response to demand.
Although we have provided a brief overview here, datacenter efficiency is a topic
that is beyond the scope of this book. For more details, consult Barroso and Hölzle [18]
and Hamilton [67], who provide detailed cost models for typical modern datacenters.
However, even factoring in operational costs, evidence suggests that scaling out remains
more attractive than scaling up.
Assume failures are common. At warehouse scale, failures are not only inevitable,
but commonplace. A simple calculation suffices to demonstrate: let us suppose that a
cluster is built from reliable machines with a mean-time between failures (MTBF) of
1000 days (about three years). Even with these reliable servers, a 10,000-server cluster
would still experience roughly 10 failures a day. For the sake of argument, let us suppose
that a MTBF of 10,000 days (about thirty years) were achievable at realistic costs (which
is unlikely). Even then, a 10,000-server cluster would still experience one failure daily.
This means that any large-scale service that is distributed across a large cluster (either
a user-facing application or a computing platform like MapReduce) must cope with
hardware failures as an intrinsic aspect of its operation [66]. That is, a server may fail at
any time, without notice. For example, in large clusters disk failures are common [123]
and RAM experiences more errors than one might expect [135]. Datacenters suffer
from both planned outages (e.g., system maintenance and hardware upgrades) and
unexpected outages (e.g., power failure, connectivity loss, etc.).
A well-designed, fault-tolerant service must cope with failures up to a point with-
out impacting the quality of service—failures should not result in inconsistencies or in-
determinism from the user perspective. As servers go down, other cluster nodes should
seamlessly step in to handle the load, and overall performance should gracefully degrade
as server failures pile up. Just as important, a broken server that has been repaired
12 CHAPTER 1. INTRODUCTION
should be able to seamlessly rejoin the service without manual reconfiguration by the
administrator. Mature implementations of the MapReduce programming model are able
to robustly cope with failures through a number of mechanisms such as automatic task
restarts on different cluster nodes.
take advantage of the aggregate bandwidth of many disks in a cluster. Many aspects of
MapReduce’s design explicitly trade latency for throughput.
Seamless scalability. For data-intensive processing, it goes without saying that scal-
able algorithms are highly desirable. As an aspiration, let us sketch the behavior of an
ideal algorithm. We can define scalability along at least two dimensions.12 First, in terms
of data: given twice the amount of data, the same algorithm should take at most twice
as long to run, all else being equal. Second, in terms of resources: given a cluster twice
12 See also DeWitt and Gray [50] for slightly different definitions in terms of speedup and scaleup.
14 CHAPTER 1. INTRODUCTION
the size, the same algorithm should take no more than half as long to run. Furthermore,
an ideal algorithm would maintain these desirable scaling characteristics across a wide
range of settings: on data ranging from gigabytes to petabytes, on clusters consisting
of a few to a few thousand machines. Finally, the ideal algorithm would exhibit these
desired behaviors without requiring any modifications whatsoever, not even tuning of
parameters.
Other than for embarrassingly parallel problems, algorithms with the character-
istics sketched above are, of course, unobtainable. One of the fundamental assertions
in Fred Brook’s classic The Mythical Man-Month [28] is that adding programmers to a
project behind schedule will only make it fall further behind. This is because complex
tasks cannot be chopped into smaller pieces and allocated in a linear fashion, and is
often illustrated with a cute quote: “nine women cannot have a baby in one month”.
Although Brook’s observations are primarily about software engineers and the soft-
ware development process, the same is also true of algorithms: increasing the degree
of parallelization also increases communication costs. The algorithm designer is faced
with diminishing returns, and beyond a certain point, greater efficiencies gained by
parallelization are entirely offset by increased communication requirements.
Nevertheless, these fundamental limitations shouldn’t prevent us from at least
striving for the unobtainable. The truth is that most current algorithms are far from
the ideal. In the domain of text processing, for example, most algorithms today assume
that data fits in memory on a single machine. For the most part, this is a fair assumption.
But what happens when the amount of data doubles in the near future, and then doubles
again shortly thereafter? Simply buying more memory is not a viable solution, as the
amount of data is growing faster than the price of memory is falling. Furthermore, the
price of a machine does not scale linearly with the amount of available memory beyond
a certain point (once again, the scaling “up” vs. scaling “out” argument). Quite simply,
algorithms that require holding intermediate data in memory on a single machine will
simply break on sufficiently-large datasets—moving from a single machine to a cluster
architecture requires fundamentally different algorithms (and reimplementations).
Perhaps the most exciting aspect of MapReduce is that it represents a small step
toward algorithms that behave in the ideal manner discussed above. Recall that the
programming model maintains a clear separation between what computations need to
occur with how those computations are actually orchestrated on a cluster. As a result,
a MapReduce algorithm remains fixed, and it is the responsibility of the execution
framework to execute the algorithm. Amazingly, the MapReduce programming model
is simple enough that it is actually possible, in many circumstances, to approach the
ideal scaling characteristics discussed above. We introduce the idea of the “tradeable
machine hour”, as a play on Brook’s classic title. If running an algorithm on a particular
dataset takes 100 machine hours, then we should be able to finish in an hour on a cluster
1.3. WHY IS THIS DIFFERENT? 15
of 100 machines, or use a cluster of 10 machines to complete the same task in ten hours. 13
With MapReduce, this isn’t so far from the truth, at least for some applications.
and the open source community. This in turn created more demand: when organiza-
tions learned about the availability of effective data analysis tools for large datasets,
they began instrumenting various business processes to gather even more data—driven
by the belief that more data leads to deeper insights and greater competitive advantages.
Today, not only are large-data problems ubiquitous, but technological solutions for ad-
dressing them are widely accessible. Anyone can download the open source Hadoop
implementation of MapReduce, pay a modest fee to rent a cluster from a utility cloud
provider, and be happily processing terabytes upon terabytes of data within the week.
Finally, the computer scientists are right—the age of parallel computing has begun,
both in terms of multiple cores in a chip and multiple machines in a cluster (each of
which often has multiple cores).
Why is MapReduce important? In practical terms, it provides a very effective tool
for tackling large-data problems. But beyond that, MapReduce is important in how it
has changed the way we organize computations at a massive scale. MapReduce repre-
sents the first widely-adopted step away from the von Neumann model that has served
as the foundation of computer science over the last half plus century. Valiant called this
a bridging model [148], a conceptual bridge between the physical implementation of a
machine and the software that is to be executed on that machine. Until recently, the
von Neumann model has served us well: Hardware designers focused on efficient imple-
mentations of the von Neumann model and didn’t have to think much about the actual
software that would run on the machines. Similarly, the software industry developed
software targeted at the model without worrying about the hardware details. The result
was extraordinary growth: chip designers churned out successive generations of increas-
ingly powerful processors, and software engineers were able to develop applications in
high-level languages that exploited those processors.
Today, however, the von Neumann model isn’t sufficient anymore: we can’t treat
a multi-core processor or a large cluster as an agglomeration of many von Neumann
machine instances communicating over some interconnect. Such a view places too much
burden on the software developer to effectively take advantage of available computa-
tional resources—it simply is the wrong level of abstraction. MapReduce can be viewed
as the first breakthrough in the quest for new abstractions that allow us to organize
computations, not over individual machines, but over entire clusters. As Barroso puts
it, the datacenter is the computer [18, 119].
To be fair, MapReduce is certainly not the first model of parallel computation
that has been proposed. The most prevalent model in theoretical computer science,
which dates back several decades, is the PRAM [77, 60].15 In the model, an arbitrary
number of processors, sharing an unboundedly large memory, operate synchronously on
a shared input to produce some output. Other models include LogP [43] and BSP [148].
15 More than a theoretical model, the PRAM has been recently prototyped in hardware [153].
1.4. WHAT THIS BOOK IS NOT 17
For reasons that are beyond the scope of this book, none of these previous models have
enjoyed the success that MapReduce has in terms of adoption and in terms of impact
on the daily lives of millions of users.16
MapReduce is the most successful abstraction over large-scale computational re-
sources we have seen to date. However, as anyone who has taken an introductory
computer science course knows, abstractions manage complexity by hiding details and
presenting well-defined behaviors to users of those abstractions. They, inevitably, are
imperfect—making certain tasks easier but others more difficult, and sometimes, im-
possible (in the case where the detail suppressed by the abstraction is exactly what
the user cares about). This critique applies to MapReduce: it makes certain large-data
problems easier, but suffers from limitations as well. This means that MapReduce is
not the final word, but rather the first in a new class of programming models that will
allow us to more effectively organize computations at a massive scale.
So if MapReduce is only the beginning, what’s next beyond MapReduce? We’re
getting ahead of ourselves, as we can’t meaningfully answer this question before thor-
oughly understanding what MapReduce can and cannot do well. This is exactly the
purpose of this book: let us now begin our exploration.
16 Nevertheless,it is important to understand the relationship between MapReduce and existing models so that we
can bring to bear accumulated knowledge about parallel algorithms; for example, Karloff et al. [82] demonstrated
that a large class of PRAM algorithms can be efficiently simulated via MapReduce.
18
CHAPTER 2
MapReduce Basics
The only feasible approach to tackling large-data problems today is to divide and con-
quer, a fundamental concept in computer science that is introduced very early in typical
undergraduate curricula. The basic idea is to partition a large problem into smaller sub-
problems. To the extent that the sub-problems are independent [5], they can be tackled
in parallel by different workers—threads in a processor core, cores in a multi-core pro-
cessor, multiple processors in a machine, or many machines in a cluster. Intermediate
results from each individual worker are then combined to yield the final output.1
The general principles behind divide-and-conquer algorithms are broadly applica-
ble to a wide range of problems in many different application domains. However, the
details of their implementations are varied and complex. For example, the following are
just some of the issues that need to be addressed:
• How do we break up a large problem into smaller tasks? More specifically, how do
we decompose the problem so that the smaller tasks can be executed in parallel?
• How do we ensure that the workers get the data they need?
• How do we share partial results from one worker that is needed by another?
• How do we accomplish all of the above in the face of software errors and hardware
faults?
OpenMP for shared memory parallelism, or libraries implementing the Message Pass-
2
ing Interface (MPI) for cluster-level parallelism,3 provide logical abstractions that hide
details of operating system synchronization and communications primitives. However,
even with these extensions, developers are still burdened to keep track of how resources
are made available to workers. Additionally, these frameworks are mostly designed to
tackle processor-intensive problems and have only rudimentary support for dealing with
very large amounts of input data. When using existing parallel computing approaches
for large-data computation, the programmer must devote a significant amount of at-
tention to low-level system details, which detracts from higher-level problem solving.
One of the most significant advantages of MapReduce is that it provides an ab-
straction that hides many system-level details from the programmer. Therefore, a devel-
oper can focus on what computations need to be performed, as opposed to how those
computations are actually carried out or how to get the data to the processes that
depend on them. Like OpenMP and MPI, MapReduce provides a means to distribute
computation without burdening the programmer with the details of distributed com-
puting (but at a different level of granularity). However, organizing and coordinating
large amounts of computation is only part of the challenge. Large-data processing by
definition requires bringing data and code together for computation to occur—no small
feat for datasets that are terabytes and perhaps petabytes in size! MapReduce addresses
this challenge by providing a simple abstraction for the developer, transparently han-
dling most of the details behind the scenes in a scalable, robust, and efficient manner.
As we mentioned in Chapter 1, instead of moving large amounts of data around, it is far
more efficient, if possible, to move the code to the data. This is operationally realized
by spreading data across the local disks of nodes in a cluster and running processes
on nodes that hold the data. The complex task of managing storage in such a process-
ing environment is typically handled by a distributed file system that sits underneath
MapReduce.
This chapter introduces the MapReduce programming model and the underlying
distributed file system. We start in Section 2.1 with an overview of functional program-
ming, from which MapReduce draws its inspiration. Section 2.2 introduces the basic
programming model, focusing on mappers and reducers. Section 2.3 discusses the role
of the execution framework in actually running MapReduce programs (called jobs).
Section 2.4 fills in additional details by introducing partitioners and combiners, which
provide greater control over data flow. MapReduce would not be practical without a
tightly-integrated distributed file system that manages the data being processed; Sec-
tion 2.5 covers this in detail. Tying everything together, a complete cluster architecture
is described in Section 2.6 before the chapter ends with a summary.
2 http://www.openmp.org/
3 http://www.mcs.anl.gov/mpi/
20 CHAPTER 2. MAPREDUCE BASICS
f f f f f
g g g g g
Figure 2.1: Illustration of map and fold, two higher-order functions commonly used together
in functional programming: map takes a function f and applies it to every element in a list,
while fold iteratively applies a function g to aggregate results.
A α B β C γ D δ E ε F ζ
a 1 b 2 c 3 c 6 a 5 c 2 b 7 c 8
a 1 5 b 2 7 c 2 9 8
X 5 Y 7 Z 9
Figure 2.2: Simplified view of MapReduce. Mappers are applied to all input key-value pairs,
which generate an arbitrary number of intermediate key-value pairs. Reducers are applied to
all values associated with the same key. Between the map and reduce phases lies a barrier that
involves a large distributed sort and group by.
1: class Mapper
2: method Map(docid a, doc d)
3: for all term t ∈ doc d do
4: Emit(term t, count 1)
1: class Reducer
2: method Reduce(term t, counts [c1 , c2 , . . .])
3: sum ← 0
4: for all count c ∈ counts [c1 , c2 , . . .] do
5: sum ← sum + c
6: Emit(term t, count sum)
Figure 2.3: Pseudo-code for the word count algorithm in MapReduce. The mapper emits an
intermediate key-value pair for each word in a document. The reducer sums up all counts for
each word.
24 CHAPTER 2. MAPREDUCE BASICS
distribution over words in a collection). Input key-values pairs take the form of (docid,
doc) pairs stored on the distributed file system, where the former is a unique identifier
for the document, and the latter is the text of the document itself. The mapper takes
an input key-value pair, tokenizes the document, and emits an intermediate key-value
pair for every word: the word itself serves as the key, and the integer one serves as the
value (denoting that we’ve seen the word once). The MapReduce execution framework
guarantees that all values associated with the same key are brought together in the
reducer. Therefore, in our word count algorithm, we simply need to sum up all counts
(ones) associated with each word. The reducer does exactly this, and emits final key-
value pairs with the word as the key, and the count as the value. Final output is written
to the distributed file system, one file per reducer. Words within each file will be sorted
by alphabetical order, and each file will contain roughly the same number of words. The
partitioner, which we discuss later in Section 2.4, controls the assignment of words to
reducers. The output can be examined by the programmer or used as input to another
MapReduce program.
There are some differences between the Hadoop implementation of MapReduce
and Google’s implementation.9 In Hadoop, the reducer is presented with a key and an
iterator over all values associated with the particular key. The values are arbitrarily
ordered. Google’s implementation allows the programmer to specify a secondary sort
key for ordering the values (if desired)—in which case values associated with each key
would be presented to the developer’s reduce code in sorted order. Later in Section 3.4
we discuss how to overcome this limitation in Hadoop to perform secondary sorting.
Another difference: in Google’s implementation the programmer is not allowed to change
the key in the reducer. That is, the reducer output key must be exactly the same as the
reducer input key. In Hadoop, there is no such restriction, and the reducer can emit an
arbitrary number of output key-value pairs (with different keys).
To provide a bit more implementation detail: pseudo-code provided in this book
roughly mirrors how MapReduce programs are written in Hadoop. Mappers and reduc-
ers are objects that implement the Map and Reduce methods, respectively. In Hadoop,
a mapper object is initialized for each map task (associated with a particular sequence
of key-value pairs called an input split) and the Map method is called on each key-value
pair by the execution framework. In configuring a MapReduce job, the programmer pro-
vides a hint on the number of map tasks to run, but the execution framework (see next
section) makes the final determination based on the physical layout of the data (more
details in Section 2.5 and Section 2.6). The situation is similar for the reduce phase:
a reducer object is initialized for each reduce task, and the Reduce method is called
once per intermediate key. In contrast with the number of map tasks, the programmer
can precisely specify the number of reduce tasks. We will return to discuss the details
mappers and reducers has the effect of regrouping and resorting the input data (which
is sometimes useful).
Although in the most common case, input to a MapReduce job comes from data
stored on the distributed file system and output is written back to the distributed file
system, any other system that satisfies the proper abstractions can serve as a data source
or sink. With Google’s MapReduce implementation, BigTable [34], a sparse, distributed,
persistent multidimensional sorted map, is frequently used as a source of input and as
a store of MapReduce output. HBase is an open-source BigTable clone and has similar
capabilities. Also, Hadoop has been integrated with existing MPP (massively parallel
processing) relational databases, which allows a programmer to write MapReduce jobs
over database rows and dump output into a new database table. Finally, in some cases
MapReduce jobs may not consume any input at all (e.g., computing π) or may only
consume a small amount of data (e.g., input parameters to many instances of processor-
intensive simulations running in parallel).
Scheduling. Each MapReduce job is divided into smaller units called tasks (see Sec-
tion 2.6 for more details). For example, a map task may be responsible for processing
a certain block of input key-value pairs (called an input split in Hadoop); similarly, a
reduce task may handle a portion of the intermediate key space. It is not uncommon
for MapReduce jobs to have thousands of individual tasks that need to be assigned to
nodes in the cluster. In large jobs, the total number of tasks may exceed the number of
tasks that can be run on the cluster concurrently, making it necessary for the scheduler
to maintain some sort of a task queue and to track the progress of running tasks so
that waiting tasks can be assigned to nodes as they become available. Another aspect
of scheduling involves coordination among tasks belonging to different jobs (e.g., from
different users). How can a large, shared resource support several users simultaneously
in a predictable, transparent, policy-driven fashion? There has been some recent work
along these lines in the context of Hadoop [131, 160].
2.3. THE EXECUTION FRAMEWORK 27
Data/code co-location. The phrase data distribution is misleading, since one of the
key ideas behind MapReduce is to move the code, not the data. However, the more
general point remains—in order for computation to occur, we need to somehow feed
data to the code. In MapReduce, this issue is inexplicably intertwined with scheduling
and relies heavily on the design of the underlying distributed file system.11 To achieve
data locality, the scheduler starts tasks on the node that holds a particular block of data
(i.e., on its local drive) needed by the task. This has the effect of moving code to the
data. If this is not possible (e.g., a node is already running too many tasks), new tasks
will be started elsewhere, and the necessary data will be streamed over the network.
An important optimization here is to prefer nodes that are on the same rack in the
datacenter as the node holding the relevant data block, since inter-rack bandwidth is
significantly less than intra-rack bandwidth.
sort involving all the nodes that executed map tasks and all the nodes that will execute
reduce tasks. This necessarily involves copying intermediate data over the network, and
therefore the process is commonly known as “shuffle and sort”. A MapReduce job with
m mappers and r reducers involves up to m × r distinct copy operations, since each
mapper may have intermediate output going to every reducer.
Note that the reduce computation cannot start until all the mappers have fin-
ished emitting key-value pairs and all intermediate key-value pairs have been shuffled
and sorted, since the execution framework cannot otherwise guarantee that all values
associated with the same key have been gathered. This is an important departure from
functional programming: in a fold operation, the aggregation function g is a function of
the intermediate value and the next item in the list—which means that values can be
lazily generated and aggregation can begin as soon as values are available. In contrast,
the reducer in MapReduce receives all values associated with the same key at once.
However, it is possible to start copying intermediate key-value pairs over the network
to the nodes running the reducers as soon as each mapper finishes—this is a common
optimization and implemented in Hadoop.
Error and fault handling. The MapReduce execution framework must accomplish
all the tasks above in an environment where errors and faults are the norm, not the
exception. Since MapReduce was explicitly designed around low-end commodity servers,
the runtime must be especially resilient. In large clusters, disk failures are common [123]
and RAM experiences more errors than one might expect [135]. Datacenters suffer
from both planned outages (e.g., system maintenance and hardware upgrades) and
unexpected outages (e.g., power failure, connectivity loss, etc.).
And that’s just hardware. No software is bug free—exceptions must be appropri-
ately trapped, logged, and recovered from. Large-data problems have a penchant for
uncovering obscure corner cases in code that is otherwise thought to be bug-free. Fur-
thermore, any sufficiently large dataset will contain corrupted data or records that are
mangled beyond a programmer’s imagination—resulting in errors that one would never
think to check for or trap. The MapReduce execution framework must thrive in this
hostile environment.
simplest partitioner involves computing the hash value of the key and then taking the
mod of that value with the number of reducers. This assigns approximately the same
number of keys to each reducer (dependent on the quality of the hash function). Note,
however, that the partitioner only considers the key and ignores the value—therefore, a
roughly-even partitioning of the key space may nevertheless yield large differences in the
number of key-values pairs sent to each reducer (since different keys may have different
numbers of associated values). This imbalance in the amount of data associated with
each key is relatively common in many text processing applications due to the Zipfian
distribution of word occurrences.
Combiners are an optimization in MapReduce that allow for local aggregation
before the shuffle and sort phase. We can motivate the need for combiners by considering
the word count algorithm in Figure 2.3, which emits a key-value pair for each word
in the collection. Furthermore, all these key-value pairs need to be copied across the
network, and so the amount of intermediate data will be larger than the input collection
itself. This is clearly inefficient. One solution is to perform local aggregation on the
output of each mapper, i.e., to compute a local count for a word over all the documents
processed by the mapper. With this modification (assuming the maximum amount of
local aggregation possible), the number of intermediate key-value pairs will be at most
the number of unique words in the collection times the number of mappers (and typically
far smaller because each mapper may not encounter every word).
The combiner in MapReduce supports such an optimization. One can think of
combiners as “mini-reducers” that take place on the output of the mappers, prior to the
shuffle and sort phase. Each combiner operates in isolation and therefore does not have
access to intermediate output from other mappers. The combiner is provided keys and
values associated with each key (the same types as the mapper output keys and values).
Critically, one cannot assume that a combiner will have the opportunity to process all
values associated with the same key. The combiner can emit any number of key-value
pairs, but the keys and values must be of the same type as the mapper output (same as
the reducer input).12 In cases where an operation is both associative and commutative
(e.g., addition or multiplication), reducers can directly serve as combiners. In general,
however, reducers and combiners are not interchangeable.
In many cases, proper use of combiners can spell the difference between an imprac-
tical algorithm and an efficient algorithm. This topic will be discussed in Section 3.1,
which focuses on various techniques for local aggregation. It suffices to say for now that
12 A note on the implementation of combiners in Hadoop: by default, the execution framework reserves the right
to use combiners at its discretion. In reality, this means that a combiner may be invoked zero, one, or multiple
times. In addition, combiners in Hadoop may actually be invoked in the reduce phase, i.e., after key-value pairs
have been copied over to the reducer, but before the user reducer code runs. As a result, combiners must be
carefully written so that they can be executed in these different environments. Section 3.1.2 discusses this in
more detail.
30 CHAPTER 2. MAPREDUCE BASICS
A α B β C γ D δ E ε F ζ
mapper
pp mapper
pp mapper
pp mapper
pp
a 1 b 2 c 3 c 6 a 5 c 2 b 7 c 8
a 1 b 2 c 9 a 5 c 2 b 7 c 8
partitioner
p partitioner
p partitioner
p partitioner
p
a 1 5 b 2 7 c 2 9 8
X 5 Y 7 Z 9
Figure 2.4: Complete view of MapReduce, illustrating combiners and partitioners in addi-
tion to mappers and reducers. Combiners can be viewed as “mini-reducers” in the map phase.
Partitioners determine which reducer is responsible for a particular key.
a combiner can significantly reduce the amount of data that needs to be copied over
the network, resulting in much faster algorithms.
The complete MapReduce model is shown in Figure 2.4. Output of the mappers
are processed by the combiners, which perform local aggregation to cut down on the
number of intermediate key-value pairs. The partitioner determines which reducer will
be responsible for processing a particular key, and the execution framework uses this
information to copy the data to the right location during the shuffle and sort phase.13
Therefore, a complete MapReduce job consists of code for the mapper, reducer, com-
biner, and partitioner, along with job configuration parameters. The execution frame-
work handles everything else.
13 In Hadoop, partitioners are actually executed before combiners, so while Figure 2.4 is conceptually accurate,
it doesn’t precisely describe the Hadoop implementation.
2.5. THE DISTRIBUTED FILE SYSTEM 31
HDFS namenode
Application /foo/bar
(file name, block id)
File namespace block 3df2
HDFS Client
(block id, block location)
instructions to datanode
datanode state
(block id, byte range)
HDFS datanode HDFS datanode
block data
Linux file system Linux file system
… …
Figure 2.5: The architecture of HDFS. The namenode (master) is responsible for maintaining
the file namespace and directing clients to datanodes (slaves) that actually hold data blocks
containing user data.
data blocks. In GFS, the master is called the GFS master, and the slaves are called
GFS chunkservers. In Hadoop, the same roles are filled by the namenode and datan-
odes, respectively.15 This book adopts the Hadoop terminology, although for most basic
file operations GFS and HDFS work much the same way. The architecture of HDFS is
shown in Figure 2.5, redrawn from a similar diagram describing GFS [57].
In HDFS, an application client wishing to read a file (or a portion thereof) must
first contact the namenode to determine where the actual data is stored. In response
to the client request, the namenode returns the relevant block id and the location
where the block is held (i.e., which datanode). The client then contacts the datanode to
retrieve the data. Blocks are themselves stored on standard single-machine file systems,
so HDFS lies on top of the standard OS stack (e.g., Linux). An important feature of
the design is that data is never moved through the namenode. Instead, all data transfer
occurs directly between clients and datanodes; communications with the namenode only
involves transfer of metadata.
By default, HDFS stores three separate copies of each data block to ensure both
reliability, availability, and performance. In large clusters, the three replicas are spread
across different physical racks, so HDFS is resilient towards two common failure sce-
narios: individual datanode crashes and failures in networking equipment that bring
an entire rack offline. Replicating blocks across physical machines also increases oppor-
15 To be precise, namenode and datanode may refer to physical machines in a cluster, or they may refer to daemons
running on those machines providing the relevant services.
2.5. THE DISTRIBUTED FILE SYSTEM 33
tunities to co-locate data and processing in the scheduling of MapReduce jobs, since
multiple copies yield more opportunities to exploit locality. The namenode is in periodic
communication with the datanodes to ensure proper replication of all the blocks: if there
aren’t enough replicas (e.g., due to disk or machine failures or to connectivity losses
due to networking equipment failures), the namenode directs the creation of additional
copies;16 if there are too many replicas (e.g., a repaired node rejoins the cluster), extra
copies are discarded.
To create a new file and write data to HDFS, the application client first contacts
the namenode, which updates the file namespace after checking permissions and making
sure the file doesn’t already exist. The namenode allocates a new block on a suitable
datanode, and the application is directed to stream data directly to it. From the initial
datanode, data is further propagated to additional replicas. In the most recent release of
Hadoop as of this writing (release 0.20.2), files are immutable—they cannot be modified
after creation. There are current plans to officially support file appends in the near
future, which is a feature already present in GFS.
In summary, the HDFS namenode has the following responsibilities:
• Maintaining overall health of the file system. The namenode is in periodic contact
with the datanodes via heartbeat messages to ensure the integrity of the system.
If the namenode observes that a data block is under-replicated (fewer copies are
stored on datanodes than the desired replication factor), it will direct the creation
of new replicas. Finally, the namenode is also responsible for rebalancing the file
system.17 During the course of normal operations, certain datanodes may end up
holding more blocks than others; rebalancing involves moving blocks from datan-
odes with more blocks to datanodes with fewer blocks. This leads to better load
balancing and more even disk utilization.
16 Note that the namenode coordinates the replication process, but data transfer occurs directly from datanode
to datanode.
17 In Hadoop, this is a manually-invoked process.
34 CHAPTER 2. MAPREDUCE BASICS
Since GFS and HDFS were specifically designed to support Google’s proprietary and
the open-source implementation of MapReduce, respectively, they were designed with
a number of assumptions about the operational environment, which in turn influenced
the design of the systems. Understanding these choices is critical to designing effective
MapReduce algorithms:
• The file system stores a relatively modest number of large files. The definition of
“modest” varies by the size of the deployment, but in HDFS multi-gigabyte files
are common (and even encouraged). There are several reasons why lots of small
files are to be avoided. Since the namenode must hold all file metadata in memory,
this presents an upper bound on both the number of files and blocks that can
be supported.18 Large multi-block files represent a more efficient use of namenode
memory than many single-block files (each of which consumes less space than a
single block size). In addition, mappers in a MapReduce job use individual files as
a basic unit for splitting input data. At present, there is no default mechanism in
Hadoop that allows a mapper to process multiple files. As a result, mapping over
many small files will yield as many map tasks as there are files. This results in
two potential problems: first, the startup costs of mappers may become significant
compared to the time spent actually processing input key-value pairs; second, this
may result in an excessive amount of across-the-network copy operations during
the “shuffle and sort” phase (recall that a MapReduce job with m mappers and r
reducers involves up to m × r distinct copy operations).
• Workloads are batch oriented, dominated by long streaming reads and large se-
quential writes. As a result, high sustained bandwidth is more important than low
latency. This exactly describes the nature of MapReduce jobs, which are batch
operations on large amounts of data. Due to the common-case workload, both
HDFS and GFS do not implement any form of data caching.19
• Applications are aware of the characteristics of the distributed file system. Neither
HDFS nor GFS present a general POSIX-compliant API, but rather support only
a subset of possible file operations. This simplifies the design of the distributed
file system, and in essence pushes part of the data management onto the end
application. One rationale for this decision is that each application knows best
how to handle data specific to that application, for example, in terms of resolving
inconsistent states and optimizing the layout of data structures.
18 According to Dhruba Borthakur in a post to the Hadoop mailing list on 6/8/2008, each block in HDFS occupies
about 150 bytes of memory on the namenode.
19 However, since the distributed file system is built on top of a standard operating system such as Linux, there
t kt k
tasktracker t kt k
tasktracker t kt k
tasktracker
… … …
slave node slave node slave node
Figure 2.6: Architecture of a complete Hadoop cluster, which consists of three separate compo-
nents: the HDFS master (called the namenode), the job submission node (called the jobtracker),
and many slave nodes (three shown here). Each of the slave nodes runs a tasktracker for exe-
cuting map and reduce tasks and a datanode daemon for serving HDFS data.
pairs, called an input split in Hadoop. Input splits are computed automatically and the
execution framework strives to align them to HDFS block boundaries so that each map
task is associated with a single data block. In scheduling map tasks, the jobtracker tries
to take advantage of data locality—if possible, map tasks are scheduled on the slave
node that holds the input split, so that the mapper will be processing local data. The
alignment of input splits with HDFS block boundaries simplifies task scheduling. If it
is not possible to run a map task on local data, it becomes necessary to stream input
key-value pairs across the network. Since large clusters are organized into racks, with
far greater intra-rack bandwidth than inter-rack bandwidth, the execution framework
strives to at least place map tasks on a rack which has a copy of the data block.
Although conceptually in MapReduce one can think of the mapper being applied
to all input key-value pairs and the reducer being applied to all values associated with
the same key, actual job execution is a bit more complex. In Hadoop, mappers are Java
objects with a Map method (among others). A mapper object is instantiated for every
map task by the tasktracker. The life-cycle of this object begins with instantiation,
where a hook is provided in the API to run programmer-specified code. This means
that mappers can read in “side data”, providing an opportunity to load state, static
data sources, dictionaries, etc. After initialization, the Map method is called (by the
execution framework) on all key-value pairs in the input split. Since these method
calls occur in the context of the same Java object, it is possible to preserve state across
multiple input key-value pairs within the same map task—this is an important property
to exploit in the design of MapReduce algorithms, as we will see in the next chapter.
After all key-value pairs in the input split have been processed, the mapper object
provides an opportunity to run programmer-specified termination code. This, too, will
be important in the design of MapReduce algorithms.
The actual execution of reducers is similar to that of the mappers. Each re-
ducer object is instantiated for every reduce task. The Hadoop API provides hooks for
programmer-specified initialization and termination code. After initialization, for each
intermediate key in the partition (defined by the partitioner), the execution framework
repeatedly calls the Reduce method with an intermediate key and an iterator over
all values associated with that key. The programming model also guarantees that in-
termediate keys will be presented to the Reduce method in sorted order. Since this
occurs in the context of a single object, it is possible to preserve state across multiple
intermediate keys (and associated values) within a single reduce task. Once again, this
property is critical in the design of MapReduce algorithms and will be discussed in the
next chapter.
38 CHAPTER 2. MAPREDUCE BASICS
2.7 SUMMARY
This chapter provides a basic overview of the MapReduce programming model, starting
with its roots in functional programming and continuing with a description of mappers,
reducers, partitioners, and combiners. Significant attention is also given to the underly-
ing distributed file system, which is a tightly-integrated component of the MapReduce
environment. Given this basic understanding, we now turn our attention to the design
of MapReduce algorithms.
39
CHAPTER 3
1. The ability to construct complex data structures as keys and values to store and
communicate partial results.
2. The ability to execute user-specified initialization code at the beginning of a map
or reduce task, and the ability to execute user-specified termination code at the
end of a map or reduce task.
3. The ability to preserve state in both mappers and reducers across multiple input
or intermediate keys.
4. The ability to control the sort order of intermediate keys, and therefore the order
in which a reducer will encounter particular keys.
5. The ability to control the partitioning of the key space, and therefore the set of
keys that will be encountered by a particular reducer.
It is important to realize that many algorithms cannot be easily expressed as a single
MapReduce job. One must often decompose complex algorithms into a sequence of jobs,
which requires orchestrating data so that the output of one job becomes the input to the
next. Many algorithms are iterative in nature, requiring repeated execution until some
convergence criteria—graph algorithms in Chapter 5 and expectation-maximization al-
gorithms in Chapter 6 behave in exactly this way. Often, the convergence check itself
cannot be easily expressed in MapReduce. The standard solution is an external (non-
MapReduce) program that serves as a “driver” to coordinate MapReduce iterations.
This chapter explains how various techniques to control code execution and
data flow can be applied to design algorithms in MapReduce. The focus is both on
scalability—ensuring that there are no inherent bottlenecks as algorithms are applied
to increasingly larger datasets—and efficiency—ensuring that algorithms do not need-
lessly consume resources and thereby reducing the cost of parallelization. The gold
standard, of course, is linear scalability: an algorithm running on twice the amount
of data should take only twice as long. Similarly, an algorithm running on twice the
number of nodes should only take half as long.
The chapter is organized as follows:
• Section 3.1 introduces the important concept of local aggregation in MapReduce
and strategies for designing efficient algorithms that minimize the amount of par-
tial results that need to be copied across the network. The proper use of combiners
is discussed in detail, as well as the “in-mapper combining” design pattern.
• Section 3.2 uses the example of building word co-occurrence matrices on large
text corpora to illustrate two common design patterns, which we dub “pairs” and
“stripes”. These two approaches are useful in a large class of problems that require
keeping track of joint events across a large number of observations.
3.1. LOCAL AGGREGATION 41
• Section 3.3 shows how co-occurrence counts can be converted into relative frequen-
cies using a pattern known as “order inversion”. The sequencing of computations
in the reducer can be recast as a sorting problem, where pieces of intermediate
data are sorted into exactly the order that is required to carry out a series of
computations. Often, a reducer needs to compute an aggregate statistic on a set
of elements before individual elements can be processed. Normally, this would re-
quire two passes over the data, but with the “order inversion” design pattern, the
aggregate statistic can be computed in the reducer before the individual elements
are encountered. This may seem counter-intuitive: how can we compute an aggre-
gate statistic on a set of elements before encountering elements of that set? As it
turns out, clever sorting of special key-value pairs enables exactly this.
• Section 3.4 provides a general solution to secondary sorting, which is the problem
of sorting values associated with a key in the reduce phase. We call this technique
“value-to-key conversion”.
• Section 3.5 covers the topic of performing joins on relational datasets and presents
three different approaches: reduce-side, map-side, and memory-backed joins.
Figure 3.1: Pseudo-code for the basic word count algorithm in MapReduce (repeated from
Figure 2.3).
The first technique for local aggregation is the combiner, already discussed in
Section 2.4. Combiners provide a general mechanism within the MapReduce framework
to reduce the amount of intermediate data generated by the mappers—recall that they
can be understood as “mini-reducers” that process the output of mappers. In this
example, the combiners aggregate term counts across the documents processed by each
map task. This results in a reduction in the number of intermediate key-value pairs that
need to be shuffled across the network—from the order of total number of terms in the
collection to the order of the number of unique terms in the collection.1
An improvement on the basic algorithm is shown in Figure 3.2 (the mapper is
modified but the reducer remains the same as in Figure 3.1 and therefore is not re-
peated). An associative array (i.e., Map in Java) is introduced inside the mapper to
tally up term counts within a single document: instead of emitting a key-value pair for
each term in the document, this version emits a key-value pair for each unique term in
the document. Given that some words appear frequently within a document (for exam-
ple, a document about dogs is likely to have many occurrences of the word “dog”), this
can yield substantial savings in the number of intermediate key-value pairs emitted,
especially for long documents.
1 More precisely, if the combiners take advantage of all opportunities for local aggregation, the algorithm would
generate at most m × V intermediate key-value pairs, where m is the number of mappers and V is the vo-
cabulary size (number of unique terms in the collection), since every term could have been observed in every
mapper. However, there are two additional factors to consider. Due to the Zipfian nature of term distributions,
most terms will not be observed by most mappers (for example, terms that occur only once will by definition
only be observed by one mapper). On the other hand, combiners in Hadoop are treated as optional optimiza-
tions, so there is no guarantee that the execution framework will take advantage of all opportunities for partial
aggregation.
3.1. LOCAL AGGREGATION 43
1: class Mapper
2: method Map(docid a, doc d)
3: H ← new AssociativeArray
4: for all term t ∈ doc d do
5: H{t} ← H{t} + 1 . Tally counts for entire document
6: for all term t ∈ H do
7: Emit(term t, count H{t})
Figure 3.2: Pseudo-code for the improved MapReduce word count algorithm that uses an
associative array to aggregate term counts on a per-document basis. Reducer is the same as in
Figure 3.1.
This basic idea can be taken one step further, as illustrated in the variant of the
word count algorithm in Figure 3.3 (once again, only the mapper is modified). The
workings of this algorithm critically depends on the details of how map and reduce
tasks in Hadoop are executed, discussed in Section 2.6. Recall, a (Java) mapper object
is created for each map task, which is responsible for processing a block of input key-
value pairs. Prior to processing any input key-value pairs, the mapper’s Initialize
method is called, which is an API hook for user-specified code. In this case, we initialize
an associative array for holding term counts. Since it is possible to preserve state across
multiple calls of the Map method (for each input key-value pair), we can continue
to accumulate partial term counts in the associative array across multiple documents,
and emit key-value pairs only when the mapper has processed all documents. That is,
emission of intermediate data is deferred until the Close method in the pseudo-code.
Recall that this API hook provides an opportunity to execute user-specified code after
the Map method has been applied to all input key-value pairs of the input data split
to which the map task was assigned.
With this technique, we are in essence incorporating combiner functionality di-
rectly inside the mapper. There is no need to run a separate combiner, since all op-
portunities for local aggregation are already exploited.2 This is a sufficiently common
design pattern in MapReduce that it’s worth giving it a name, “in-mapper combining”,
so that we can refer to the pattern more conveniently throughout the book. We’ll see
later on how this pattern can be applied to a variety of problems. There are two main
advantages to using this design pattern:
First, it provides control over when local aggregation occurs and how it exactly
takes place. In contrast, the semantics of the combiner is underspecified in MapReduce.
2 Leaving aside the minor complication that in Hadoop, combiners can be run in the reduce phase also (when
merging intermediate key-value pairs from different map tasks). However, in practice it makes almost no
difference either way.
44 CHAPTER 3. MAPREDUCE ALGORITHM DESIGN
1: class Mapper
2: method Initialize
3: H ← new AssociativeArray
4: method Map(docid a, doc d)
5: for all term t ∈ doc d do
6: H{t} ← H{t} + 1 . Tally counts across documents
7: method Close
8: for all term t ∈ H do
9: Emit(term t, count H{t})
Figure 3.3: Pseudo-code for the improved MapReduce word count algorithm that demon-
strates the “in-mapper combining” design pattern. Reducer is the same as in Figure 3.1.
For example, Hadoop makes no guarantees on how many times the combiner is applied,
or that it is even applied at all. The combiner is provided as a semantics-preserving
optimization to the execution framework, which has the option of using it, perhaps
multiple times, or not at all (or even in the reduce phase). In some cases (although not
in this particular example), such indeterminism is unacceptable, which is exactly why
programmers often choose to perform their own local aggregation in the mappers.
Second, in-mapper combining will typically be more efficient than using actual
combiners. One reason for this is the additional overhead associated with actually ma-
terializing the key-value pairs. Combiners reduce the amount of intermediate data that
is shuffled across the network, but don’t actually reduce the number of key-value pairs
that are emitted by the mappers in the first place. With the algorithm in Figure 3.2,
intermediate key-value pairs are still generated on a per-document basis, only to be
“compacted” by the combiners. This process involves unnecessary object creation and
destruction (garbage collection takes time), and furthermore, object serialization and
deserialization (when intermediate key-value pairs fill the in-memory buffer holding map
outputs and need to be temporarily spilled to disk). In contrast, with in-mapper com-
bining, the mappers will generate only those key-value pairs that need to be shuffled
across the network to the reducers.
There are, however, drawbacks to the in-mapper combining pattern. First, it
breaks the functional programming underpinnings of MapReduce, since state is be-
ing preserved across multiple input key-value pairs. Ultimately, this isn’t a big deal,
since pragmatic concerns for efficiency often trump theoretical “purity”, but there are
practical consequences as well. Preserving state across multiple input instances means
that algorithmic behavior may depend on the order in which input key-value pairs are
encountered. This creates the potential for ordering-dependent bugs, which are difficult
to debug on large datasets in the general case (although the correctness of in-mapper
3.1. LOCAL AGGREGATION 45
combining for word count is easy to demonstrate). Second, there is a fundamental scala-
bility bottleneck associated with the in-mapper combining pattern. It critically depends
on having sufficient memory to store intermediate results until the mapper has com-
pletely processed all key-value pairs in an input split. In the word count example, the
memory footprint is bound by the vocabulary size, since it is theoretically possible that
a mapper encounters every term in the collection. Heap’s Law, a well-known result in
information retrieval, accurately models the growth of vocabulary size as a function
of the collection size—the somewhat surprising fact is that the vocabulary size never
stops growing.3 Therefore, the algorithm in Figure 3.3 will scale only up to a point,
beyond which the associative array holding the partial term counts will no longer fit in
memory.4
One common solution to limiting memory usage when using the in-mapper com-
bining technique is to “block” input key-value pairs and “flush” in-memory data struc-
tures periodically. The idea is simple: instead of emitting intermediate data only after
every key-value pair has been processed, emit partial results after processing every n
key-value pairs. This is straightforwardly implemented with a counter variable that
keeps track of the number of input key-value pairs that have been processed. As an
alternative, the mapper could keep track of its own memory footprint and flush inter-
mediate key-value pairs once memory usage has crossed a certain threshold. In both
approaches, either the block size or the memory usage threshold needs to be determined
empirically: with too large a value, the mapper may run out of memory, but with too
small a value, opportunities for local aggregation may be lost. Furthermore, in Hadoop
physical memory is split between multiple tasks that may be running on a node con-
currently; these tasks are all competing for finite resources, but since the tasks are not
aware of each other, it is difficult to coordinate resource consumption effectively. In
practice, however, one often encounters diminishing returns in performance gains with
increasing buffer sizes, such that it is not worth the effort to search for an optimal buffer
size (personal communication, Jeff Dean).
In MapReduce algorithms, the extent to which efficiency can be increased through
local aggregation depends on the size of the intermediate key space, the distribution of
keys themselves, and the number of key-value pairs that are emitted by each individual
map task. Opportunities for aggregation, after all, come from having multiple values
associated with the same key (whether one uses combiners or employs the in-mapper
combining pattern). In the word count example, local aggregation is effective because
3 In more detail, Heap’s Law relates the vocabulary size V to the collection size as follows: V = kT b , where
T is the number of tokens in the collection. Typical values of the parameters k and b are: 30 ≤ k ≤ 100 and
b ∼ 0.5 ([101], p. 81).
4 A few more details: note what matters is that the partial term counts encountered within particular input
split fits into memory. However, as collection sizes increase, one will often want to increase the input split size
to limit the growth of the number of map tasks (in order to reduce the number of distinct copy operations
necessary to shuffle intermediate data over the network).
46 CHAPTER 3. MAPREDUCE ALGORITHM DESIGN
many words are encountered multiple times within a map task. Local aggregation is also
an effective technique for dealing with reduce stragglers (see Section 2.3) that result
from a highly-skewed (e.g., Zipfian) distribution of values associated with intermediate
keys. In our word count example, we do not filter frequently-occurring words: therefore,
without local aggregation, the reducer that’s responsible for computing the count of
‘the’ will have a lot more work to do than the typical reducer, and therefore will likely
be a straggler. With local aggregation (either combiners or in-mapper combining), we
substantially reduce the number of values associated with frequently-occurring terms,
which alleviates the reduce straggler problem.
Figure 3.4: Pseudo-code for the basic MapReduce algorithm that computes the mean of values
associated with the same key.
of values associated with the same key, and the reducer would compute the mean of
those values. As a concrete example, we know that:
In general, the mean of means of arbitrary subsets of a set of numbers is not the same
as the mean of the set of numbers. Therefore, this approach would not produce the
correct result.5
So how might we properly take advantage of combiners? An attempt is shown in
Figure 3.5. The mapper remains the same, but we have added a combiner that partially
aggregates results by computing the numeric components necessary to arrive at the
mean. The combiner receives each string and the associated list of integer values, from
which it computes the sum of those values and the number of integers encountered (i.e.,
the count). The sum and count are packaged into a pair, and emitted as the output
of the combiner, with the same string as the key. In the reducer, pairs of partial sums
and counts can be aggregated to arrive at the mean. Up until now, all keys and values
in our algorithms have been primitives (string, integers, etc.). However, there are no
prohibitions in MapReduce for more complex types,6 and, in fact, this represents a key
technique in MapReduce algorithm design that we introduced at the beginning of this
5 There is, however, one special case in which using reducers as combiners would produce the correct result: if
each combiner computed the mean of equal-size subsets of the values. However, since such fine-grained control
over the combiners is impossible in MapReduce, such a scenario is highly unlikely.
6 In Hadoop, either custom types or types defined using a library such as Protocol Buffers, Thrift, or Avro.
48 CHAPTER 3. MAPREDUCE ALGORITHM DESIGN
1: class Mapper
2: method Map(string t, integer r)
3: Emit(string t, integer r)
1: class Combiner
2: method Combine(string t, integers [r1 , r2 , . . .])
3: sum ← 0
4: cnt ← 0
5: for all integer r ∈ integers [r1 , r2 , . . .] do
6: sum ← sum + r
7: cnt ← cnt + 1
8: Emit(string t, pair (sum, cnt)) . Separate sum and count
1: class Reducer
2: method Reduce(string t, pairs [(s1 , c1 ), (s2 , c2 ) . . .])
3: sum ← 0
4: cnt ← 0
5: for all pair (s, c) ∈ pairs [(s1 , c1 ), (s2 , c2 ) . . .] do
6: sum ← sum + s
7: cnt ← cnt + c
8: ravg ← sum/cnt
9: Emit(string t, integer ravg )
Figure 3.5: Pseudo-code for an incorrect first attempt at introducing combiners to compute
the mean of values associated with each key. The mismatch between combiner input and output
key-value types violates the MapReduce programming model.
chapter. We will frequently encounter complex keys and values throughput the rest of
this book.
Unfortunately, this algorithm will not work. Recall that combiners must have the
same input and output key-value type, which also must be the same as the mapper
output type and the reducer input type. This is clearly not the case. To understand
why this restriction is necessary in the programming model, remember that combiners
are optimizations that cannot change the correctness of the algorithm. So let us remove
the combiner and see what happens: the output value type of the mapper is integer,
so the reducer expects to receive a list of integers as values. But the reducer actually
expects a list of pairs! The correctness of the algorithm is contingent on the combiner
running on the output of the mappers, and more specifically, that the combiner is run
exactly once. Recall from our previous discussion that Hadoop makes no guarantees on
3.1. LOCAL AGGREGATION 49
1: class Mapper
2: method Map(string t, integer r)
3: Emit(string t, pair (r, 1))
1: class Combiner
2: method Combine(string t, pairs [(s1 , c1 ), (s2 , c2 ) . . .])
3: sum ← 0
4: cnt ← 0
5: for all pair (s, c) ∈ pairs [(s1 , c1 ), (s2 , c2 ) . . .] do
6: sum ← sum + s
7: cnt ← cnt + c
8: Emit(string t, pair (sum, cnt))
1: class Reducer
2: method Reduce(string t, pairs [(s1 , c1 ), (s2 , c2 ) . . .])
3: sum ← 0
4: cnt ← 0
5: for all pair (s, c) ∈ pairs [(s1 , c1 ), (s2 , c2 ) . . .] do
6: sum ← sum + s
7: cnt ← cnt + c
8: ravg ← sum/cnt
9: Emit(string t, integer ravg )
Figure 3.6: Pseudo-code for a MapReduce algorithm that computes the mean of values asso-
ciated with each key. This algorithm correctly takes advantage of combiners.
how many times combiners are called; it could be zero, one, or multiple times. This
violates the MapReduce programming model.
Another stab at the algorithm is shown in Figure 3.6, and this time, the algorithm
is correct. In the mapper we emit as the value a pair consisting of the integer and
one—this corresponds to a partial count over one instance. The combiner separately
aggregates the partial sums and the partial counts (as before), and emits pairs with
updated sums and counts. The reducer is similar to the combiner, except that the
mean is computed at the end. In essence, this algorithm transforms a non-associative
operation (mean of numbers) into an associative operation (element-wise sum of a pair
of numbers, with an additional division at the very end).
Let us verify the correctness of this algorithm by repeating the previous exercise:
What would happen if no combiners were run? With no combiners, the mappers would
send pairs (as values) directly to the reducers. There would be as many intermediate
pairs as there were input key-value pairs, and each of those would consist of an integer
50 CHAPTER 3. MAPREDUCE ALGORITHM DESIGN
1: class Mapper
2: method Initialize
3: S ← new AssociativeArray
4: C ← new AssociativeArray
5: method Map(string t, integer r)
6: S{t} ← S{t} + r
7: C{t} ← C{t} + 1
8: method Close
9: for all term t ∈ S do
10: Emit(term t, pair (S{t}, C{t}))
Figure 3.7: Pseudo-code for a MapReduce algorithm that computes the mean of values asso-
ciated with each key, illustrating the in-mapper combining design pattern. Only the mapper is
shown here; the reducer is the same as in Figure 3.6
and one. The reducer would still arrive at the correct sum and count, and hence the
mean would be correct. Now add in the combiners: the algorithm would remain correct,
no matter how many times they run, since the combiners merely aggregate partial sums
and counts to pass along to the reducers. Note that although the output key-value type
of the combiner must be the same as the input key-value type of the reducer, the reducer
can emit final key-value pairs of a different type.
Finally, in Figure 3.7, we present an even more efficient algorithm that exploits the
in-mapper combining pattern. Inside the mapper, the partial sums and counts associated
with each string are held in memory across input key-value pairs. Intermediate key-value
pairs are emitted only after the entire input split has been processed; similar to before,
the value is a pair consisting of the sum and count. The reducer is exactly the same as
in Figure 3.6. Moving partial aggregation from the combiner directly into the mapper
is subjected to all the tradeoffs and caveats discussed earlier this section, but in this
case the memory footprint of the data structures for holding intermediate data is likely
to be modest, making this variant algorithm an attractive option.
published work [54, 94], this section introduces two common design patterns we have
dubbed “pairs” and “stripes” that exemplify this strategy.
As a running example, we focus on the problem of building word co-occurrence
matrices from large corpora, a common task in corpus linguistics and statistical natural
language processing. Formally, the co-occurrence matrix of a corpus is a square n × n
matrix where n is the number of unique words in the corpus (i.e., the vocabulary size). A
cell mij contains the number of times word wi co-occurs with word wj within a specific
context—a natural unit such as a sentence, paragraph, or a document, or a certain
window of m words (where m is an application-dependent parameter). Note that the
upper and lower triangles of the matrix are identical since co-occurrence is a symmetric
relation, though in the general case relations between words need not be symmetric. For
example, a co-occurrence matrix M where mij is the count of how many times word i
was immediately succeeded by word j would usually not be symmetric.
This task is quite common in text processing and provides the starting point to
many other algorithms, e.g., for computing statistics such as pointwise mutual infor-
mation [38], for unsupervised sense clustering [136], and more generally, a large body
of work in lexical semantics based on distributional profiles of words, dating back to
Firth [55] and Harris [69] in the 1950s and 1960s. The task also has applications in in-
formation retrieval (e.g., automatic thesaurus construction [137] and stemming [157]),
and other related fields such as text mining. More importantly, this problem represents
a specific instance of the task of estimating distributions of discrete joint events from a
large number of observations, a very common task in statistical natural language pro-
cessing for which there are nice MapReduce solutions. Indeed, concepts presented here
are also used in Chapter 6 when we discuss expectation-maximization algorithms.
Beyond text processing, problems in many application domains share similar char-
acteristics. For example, a large retailer might analyze point-of-sale transaction records
to identify correlated product purchases (e.g., customers who buy this tend to also buy
that), which would assist in inventory management and product placement on store
shelves. Similarly, an intelligence analyst might wish to identify associations between
re-occurring financial transactions that are otherwise unrelated, which might provide a
clue in thwarting terrorist activity. The algorithms discussed in this section could be
adapted to tackle these related problems.
It is obvious that the space requirement for the word co-occurrence problem is
O(n2 ), where n is the size of the vocabulary, which for real-world English corpora can
be hundreds of thousands of words, or even billions of words in web-scale collections.7
The computation of the word co-occurrence matrix is quite simple if the entire matrix
7 The size of the vocabulary depends on the definition of a “word” and techniques (if any) for corpus pre-
processing. One common strategy is to replace all rare words (below a certain frequency) with a “special”
token such as <UNK> (which stands for “unknown”) to model out-of-vocabulary words. Another technique
involves replacing numeric digits with #, such that 1.32 and 1.19 both map to the same token (#.##).
52 CHAPTER 3. MAPREDUCE ALGORITHM DESIGN
fits into memory—however, in the case where the matrix is too big to fit in memory,
a naı̈ve implementation on a single machine can be very slow as memory is paged to
disk. Although compression techniques can increase the size of corpora for which word
co-occurrence matrices can be constructed on a single machine, it is clear that there are
inherent scalability limitations. We describe two MapReduce algorithms for this task
that can scale to large corpora.
Pseudo-code for the first algorithm, dubbed the “pairs” approach, is shown in
Figure 3.8. As usual, document ids and the corresponding contents make up the input
key-value pairs. The mapper processes each input document and emits intermediate
key-value pairs with each co-occurring word pair as the key and the integer one (i.e.,
the count) as the value. This is straightforwardly accomplished by two nested loops:
the outer loop iterates over all words (the left element in the pair), and the inner
loop iterates over all neighbors of the first word (the right element in the pair). The
neighbors of a word can either be defined in terms of a sliding window or some other
contextual unit such as a sentence. The MapReduce execution framework guarantees
that all values associated with the same key are brought together in the reducer. Thus,
in this case the reducer simply sums up all the values associated with the same co-
occurring word pair to arrive at the absolute count of the joint event in the corpus,
which is then emitted as the final key-value pair. Each pair corresponds to a cell in the
word co-occurrence matrix. This algorithm illustrates the use of complex keys in order
to coordinate distributed computations.
An alternative approach, dubbed the “stripes” approach, is presented in Fig-
ure 3.9. Like the pairs approach, co-occurring word pairs are generated by two nested
loops. However, the major difference is that instead of emitting intermediate key-value
pairs for each co-occurring word pair, co-occurrence information is first stored in an
associative array, denoted H. The mapper emits key-value pairs with words as keys
and corresponding associative arrays as values, where each associative array encodes
the co-occurrence counts of the neighbors of a particular word (i.e., its context). The
MapReduce execution framework guarantees that all associative arrays with the same
key will be brought together in the reduce phase of processing. The reducer performs an
element-wise sum of all associative arrays with the same key, accumulating counts that
correspond to the same cell in the co-occurrence matrix. The final associative array is
emitted with the same word as the key. In contrast to the pairs approach, each final
key-value pair encodes a row in the co-occurrence matrix.
It is immediately obvious that the pairs algorithm generates an immense number
of key-value pairs compared to the stripes approach. The stripes representation is much
more compact, since with pairs the left element is repeated for every co-occurring word
pair. The stripes approach also generates fewer and shorter intermediate keys, and
therefore the execution framework has less sorting to perform. However, values in the
3.2. PAIRS AND STRIPES 53
1: class Mapper
2: method Map(docid a, doc d)
3: for all term w ∈ doc d do
4: for all term u ∈ Neighbors(w) do
5: Emit(pair (w, u), count 1) . Emit count for each co-occurrence
1: class Reducer
2: method Reduce(pair p, counts [c1 , c2 , . . .])
3: s←0
4: for all count c ∈ counts [c1 , c2 , . . .] do
5: s←s+c . Sum co-occurrence counts
6: Emit(pair p, count s)
Figure 3.8: Pseudo-code for the “pairs” approach for computing word co-occurrence matrices
from large corpora.
1: class Mapper
2: method Map(docid a, doc d)
3: for all term w ∈ doc d do
4: H ← new AssociativeArray
5: for all term u ∈ Neighbors(w) do
6: H{u} ← H{u} + 1 . Tally words co-occurring with w
7: Emit(Term w, Stripe H)
1: class Reducer
2: method Reduce(term w, stripes [H1 , H2 , H3 , . . .])
3: Hf ← new AssociativeArray
4: for all stripe H ∈ stripes [H1 , H2 , H3 , . . .] do
5: Sum(Hf , H) . Element-wise sum
6: Emit(term w, stripe Hf )
Figure 3.9: Pseudo-code for the “stripes” approach for computing word co-occurrence matrices
from large corpora.
54 CHAPTER 3. MAPREDUCE ALGORITHM DESIGN
stripes approach are more complex, and come with more serialization and deserialization
overhead than with the pairs approach.
Both algorithms can benefit from the use of combiners, since the respective oper-
ations in their reducers (addition and element-wise sum of associative arrays) are both
commutative and associative. However, combiners with the stripes approach have more
opportunities to perform local aggregation because the key space is the vocabulary—
associative arrays can be merged whenever a word is encountered multiple times by
a mapper. In contrast, the key space in the pairs approach is the cross of the vocab-
ulary with itself, which is far larger—counts can be aggregated only when the same
co-occurring word pair is observed multiple times by an individual mapper (which is
less likely than observing multiple occurrences of a word, as in the stripes case).
For both algorithms, the in-mapper combining optimization discussed in the pre-
vious section can also be applied; the modification is sufficiently straightforward that
we leave the implementation as an exercise for the reader. However, the above caveats
remain: there will be far fewer opportunities for partial aggregation in the pairs ap-
proach due to the sparsity of the intermediate key space. The sparsity of the key space
also limits the effectiveness of in-memory combining, since the mapper may run out of
memory to store partial counts before all documents are processed, necessitating some
mechanism to periodically emit key-value pairs (which further limits opportunities to
perform partial aggregation). Similarly, for the stripes approach, memory management
will also be more complex than in the simple word count example. For common terms,
the associative array may grow to be quite large, necessitating some mechanism to
periodically flush in-memory structures.
It is important to consider potential scalability bottlenecks of either algorithm.
The stripes approach makes the assumption that, at any point in time, each associative
array is small enough to fit into memory—otherwise, memory paging will significantly
impact performance. The size of the associative array is bounded by the vocabulary size,
which is itself unbounded with respect to corpus size (recall the previous discussion of
Heap’s Law). Therefore, as the sizes of corpora increase, this will become an increasingly
pressing issue—perhaps not for gigabyte-sized corpora, but certainly for terabyte-sized
and petabyte-sized corpora that will be commonplace tomorrow. The pairs approach,
on the other hand, does not suffer from this limitation, since it does not need to hold
intermediate data in memory.
Given this discussion, which approach is faster? Here, we present previously-
published results [94] that empirically answered this question. We have implemented
both algorithms in Hadoop and applied them to a corpus of 2.27 million documents
from the Associated Press Worldstream (APW) totaling 5.7 GB.8 Prior to working
8 This
was a subset of the English Gigaword corpus (version 3) distributed by the Linguistic Data Consortium
(LDC catalog number LDC2007T07).
3.2. PAIRS AND STRIPES 55
with Hadoop, the corpus was first preprocessed as follows: All XML markup was re-
moved, followed by tokenization and stopword removal using standard tools from the
Lucene search engine. All tokens were then replaced with unique integers for a more
efficient encoding. Figure 3.10 compares the running time of the pairs and stripes ap-
proach on different fractions of the corpus, with a co-occurrence window size of two.
These experiments were performed on a Hadoop cluster with 19 slave nodes, each with
two single-core processors and two disks.
Results demonstrate that the stripes approach is much faster than the pairs ap-
proach: 666 seconds (∼11 minutes) compared to 3758 seconds (∼62 minutes) for the
entire corpus (improvement by a factor of 5.7). The mappers in the pairs approach gen-
erated 2.6 billion intermediate key-value pairs totaling 31.2 GB. After the combiners,
this was reduced to 1.1 billion key-value pairs, which quantifies the amount of interme-
diate data transferred across the network. In the end, the reducers emitted a total of 142
million final key-value pairs (the number of non-zero cells in the co-occurrence matrix).
On the other hand, the mappers in the stripes approach generated 653 million interme-
diate key-value pairs totaling 48.1 GB. After the combiners, only 28.8 million key-value
pairs remained. The reducers emitted a total of 1.69 million final key-value pairs (the
number of rows in the co-occurrence matrix). As expected, the stripes approach pro-
vided more opportunities for combiners to aggregate intermediate results, thus greatly
reducing network traffic in the shuffle and sort phase. Figure 3.10 also shows that both
algorithms exhibit highly desirable scaling characteristics—linear in the amount of in-
put data. This is confirmed by a linear regression applied to the running time data,
which yields an R2 value close to one.
An additional series of experiments explored the scalability of the stripes approach
along another dimension: the size of the cluster. These experiments were made possible
by Amazon’s EC2 service, which allows users to rapidly provision clusters of varying
sizes for limited durations (for more information, refer back to our discussion of utility
computing in Section 1.1). Virtualized computational units in EC2 are called instances,
and the user is charged only for the instance-hours consumed. Figure 3.11 (left) shows
the running time of the stripes algorithm (on the same corpus, with same setup as
before), on varying cluster sizes, from 20 slave “small” instances all the way up to 80
slave “small” instances (along the x-axis). Running times are shown with solid squares.
Figure 3.11 (right) recasts the same results to illustrate scaling characteristics. The
circles plot the relative size and speedup of the EC2 experiments, with respect to the
20-instance cluster. These results show highly desirable linear scaling characteristics
(i.e., doubling the cluster size makes the job twice as fast). This is confirmed by a linear
regression with an R2 value close to one.
Viewed abstractly, the pairs and stripes algorithms represent two different ap-
proaches to counting co-occurring events from a large number of observations. This
56 CHAPTER 3. MAPREDUCE ALGORITHM DESIGN
4000
"stripes" approach
"pairs" approach
3500
2
R = 0.999
3000
Running time (seconds)
2500
2000
1500
1000
R2 = 0.992
500
0
0 20 40 60 80 100
Percentage of the APW corpus
Figure 3.10: Running time of the “pairs” and “stripes” algorithms for computing word co-
occurrence matrices on different fractions of the APW corpus. These experiments were per-
formed on a Hadoop cluster with 19 slaves, each with two single-core processors and two disks.
5000
4000 4x
Running time (seconds)
R2 = 0.997
Relative speedup
3000 3x
2000 2x
1000 1x
0
10 20 30 40 50 60 70 80 90 1x 2x 3x 4x
Size of EC2 cluster (number of slave instances) Relative size of EC2 cluster
Figure 3.11: Running time of the stripes algorithm on the APW corpus with Hadoop clusters
of different sizes from EC2 (left). Scaling characteristics (relative speedup) in terms of increasing
Hadoop cluster size (right).
3.3. COMPUTING RELATIVE FREQUENCIES 57
general description captures the gist of many algorithms in fields as diverse as text
processing, data mining, and bioinformatics. For this reason, these two design patterns
are broadly useful and frequently observed in a variety of applications.
To conclude, it is worth noting that the pairs and stripes approaches represent
endpoints along a continuum of possibilities. The pairs approach individually records
each co-occurring event, while the stripes approach records all co-occurring events
with respect a conditioning event. A middle ground might be to record a subset of
the co-occurring events with respect to a conditioning event. We might divide up the
entire vocabulary into b buckets (e.g., via hashing), so that words co-occurring with
wi would be divided into b smaller “sub-stripes”, associated with ten separate keys,
(wi , 1), (wi , 2) . . . (wi , b). This would be a reasonable solution to the memory limitations
of the stripes approach, since each of the sub-stripes would be smaller. In the case of
b = |V |, where |V | is the vocabulary size, this is equivalent to the pairs approach. In
the case of b = 1, this is equivalent to the standard stripes approach.
N (wi , wj )
f (wj |wi ) = P (3.1)
w0 N (wi , w )
0
Here, N (·, ·) indicates the number of times a particular co-occurring word pair is ob-
served in the corpus. We need the count of the joint event (word co-occurrence), divided
by what is known as the marginal (the sum of the counts of the conditioning variable
co-occurring with anything else).
Computing relative frequencies with the stripes approach is straightforward. In
the reducer, counts of all words that co-occur with the conditioning variable (wi in the
above example) are available in the associative array. Therefore, it suffices to sum all
P
those counts to arrive at the marginal (i.e., w0 N (wi , w0 )), and then divide all the joint
counts by the marginal to arrive at the relative frequency for all words. This implemen-
tation requires minimal modification to the original stripes algorithm in Figure 3.9, and
58 CHAPTER 3. MAPREDUCE ALGORITHM DESIGN
key values
P
(dog, ∗) [6327, 8514, . . .] compute marginal: w0 N (dog, w0 ) = 42908
(dog, aardvark) [2,1] f (aardvark|dog) = 3/42908
(dog, aardwolf) [1] f (aardwolf|dog) = 1/42908
...
(dog, zebra) [2,1,1,1] f (zebra|dog) = 5/42908
P
(doge, ∗) [682, . . .] compute marginal: w0 N (doge, w0 ) = 1267
...
Figure 3.12: Example of the sequence of key-value pairs presented to the reducer in the pairs
algorithm for computing relative frequencies. This illustrates the application of the order inver-
sion design pattern.
to the reducer before data that is needed later. However, we still need to compute the
marginal counts. Recall that in the basic pairs algorithm, each mapper emits a key-
value pair with the co-occurring word pair as the key. To compute relative frequencies,
we modify the mapper so that it additionally emits a “special” key of the form (wi , ∗),
with a value of one, that represents the contribution of the word pair to the marginal.
Through use of combiners, these partial marginal counts will be aggregated before be-
ing sent to the reducers. Alternatively, the in-mapper combining pattern can be used
to even more efficiently aggregate marginal counts.
In the reducer, we must make sure that the special key-value pairs representing
the partial marginal contributions are processed before the normal key-value pairs rep-
resenting the joint counts. This is accomplished by defining the sort order of the keys
so that pairs with the special symbol of the form (wi , ∗) are ordered before any other
key-value pairs where the left word is wi . In addition, as with before we must also prop-
erly define the partitioner to pay attention to only the left word in each pair. With the
data properly sequenced, the reducer can directly compute the relative frequencies.
A concrete example is shown in Figure 3.12, which lists the sequence of key-value
pairs that a reducer might encounter. First, the reducer is presented with the special key
(dog, ∗) and a number of values, each of which represents a partial marginal contribution
from the map phase (assume here either combiners or in-mapper combining, so the
values represent partially aggregated counts). The reducer accumulates these counts to
P
arrive at the marginal, w0 N (dog, w0 ). The reducer holds on to this value as it processes
subsequent keys. After (dog, ∗), the reducer will encounter a series of keys representing
joint counts; let’s say the first of these is the key (dog, aardvark). Associated with this
key will be a list of values representing partial joint counts from the map phase (two
separate values in this case). Summing these counts will yield the final joint count, i.e.,
the number of times dog and aardvark co-occur in the entire collection. At this point,
60 CHAPTER 3. MAPREDUCE ALGORITHM DESIGN
since the reducer already knows the marginal, simple arithmetic suffices to compute
the relative frequency. All subsequent joint counts are processed in exactly the same
manner. When the reducer encounters the next special key-value pair (doge, ∗), the
reducer resets its internal state and starts to accumulate the marginal all over again.
Observe that the memory requirement for this algorithm is minimal, since only the
marginal (an integer) needs to be stored. No buffering of individual co-occurring word
counts is necessary, and therefore we have eliminated the scalability bottleneck of the
previous algorithm.
This design pattern, which we call “order inversion”, occurs surprisingly often
and across applications in many domains. It is so named because through proper co-
ordination, we can access the result of a computation in the reducer (for example, an
aggregate statistic) before processing the data needed for that computation. The key
insight is to convert the sequencing of computations into a sorting problem. In most
cases, an algorithm requires data in some fixed order: by controlling how keys are sorted
and how the key space is partitioned, we can present data to the reducer in the order
necessary to perform the proper computations. This greatly cuts down on the amount
of partial results that the reducer needs to hold in memory.
To summarize, the specific application of the order inversion design pattern for
computing relative frequencies requires the following:
• Emitting a special key-value pair for each co-occurring word pair in the mapper
to capture its contribution to the marginal.
• Controlling the sort order of the intermediate key so that the key-value pairs
representing the marginal contributions are processed by the reducer before any
of the pairs representing the joint word co-occurrence counts.
• Defining a custom partitioner to ensure that all pairs with the same left word are
shuffled to the same reducer.
• Preserving state across multiple keys in the reducer to first compute the marginal
based on the special key-value pairs and then dividing the joint counts by the
marginals to arrive at the relative frequencies.
As we will see in Chapter 4, this design pattern is also used in inverted index construc-
tion to properly set compression parameters for postings lists.
what if in addition to sorting by key, we also need to sort by value? Google’s MapReduce
implementation provides built-in functionality for (optional) secondary sorting, which
guarantees that values arrive in sorted order. Hadoop, unfortunately, does not have this
capability built in.
Consider the example of sensor data from a scientific experiment: there are m
sensors each taking readings on continuous basis, where m is potentially a large number.
A dump of the sensor data might look something like the following, where rx after each
timestamp represents the actual sensor readings (unimportant for this discussion, but
may be a series of values, one or more complex records, or even raw bytes of images).
(t1 , m1 , r80521 )
(t1 , m2 , r14209 )
(t1 , m3 , r76042 )
...
(t2 , m1 , r21823 )
(t2 , m2 , r66508 )
(t2 , m3 , r98347 )
Suppose we wish to reconstruct the activity at each individual sensor over time. A
MapReduce program to accomplish this might map over the raw data and emit the
sensor id as the intermediate key, with the rest of each record as the value:
m1 → (t1 , r80521 )
This would bring all readings from the same sensor together in the reducer. However,
since MapReduce makes no guarantees about the ordering of values associated with the
same key, the sensor readings will not likely be in temporal order. The most obvious
solution is to buffer all the readings in memory and then sort by timestamp before
additional processing. However, it should be apparent by now that any in-memory
buffering of data introduces a potential scalability bottleneck. What if we are working
with a high frequency sensor or sensor readings over a long period of time? What if the
sensor readings themselves are large complex objects? This approach may not scale in
these cases—the reducer would run out of memory trying to buffer all values associated
with the same key.
This is a common problem, since in many applications we wish to first group
together data one way (e.g., by sensor id), and then sort within the groupings another
way (e.g., by time). Fortunately, there is a general purpose solution, which we call the
“value-to-key conversion” design pattern. The basic idea is to move part of the value
into the intermediate key to form a composite key, and let the MapReduce execution
framework handle the sorting. In the above example, instead of emitting the sensor id
as the key, we would emit the sensor id and the timestamp as a composite key:
(m1 , t1 ) → (r80521 )
62 CHAPTER 3. MAPREDUCE ALGORITHM DESIGN
The sensor reading itself now occupies the value. We must define the intermediate key
sort order to first sort by the sensor id (the left element in the pair) and then by the
timestamp (the right element in the pair). We must also implement a custom partitioner
so that all pairs associated with the same sensor are shuffled to the same reducer.
Properly orchestrated, the key-value pairs will be presented to the reducer in the
correct sorted order:
(m1 , t1 ) → [(r80521 )]
(m1 , t2 ) → [(r21823 )]
(m1 , t3 ) → [(r146925 )]
...
However, note that sensor readings are now split across multiple keys. The reducer will
need to preserve state and keep track of when readings associated with the current
sensor end and the next sensor begin.9
The basic tradeoff between the two approaches discussed above (buffer and in-
memory sort vs. value-to-key conversion) is where sorting is performed. One can explic-
itly implement secondary sorting in the reducer, which is likely to be faster but suffers
from a scalability bottleneck.10 With value-to-key conversion, sorting is offloaded to the
MapReduce execution framework. Note that this approach can be arbitrarily extended
to tertiary, quaternary, etc. sorting. This pattern results in many more keys for the
framework to sort, but distributed sorting is a task that the MapReduce runtime excels
at since it lies at the heart of the programming model.
within the reducer, although one would be duplicating functionality that is already present in the MapReduce
execution framework. It makes more sense to take advantage of functionality that is already present with
value-to-key conversion.
3.5. RELATIONAL JOINS 63
find that they often cannot cost-effectively scale to the crushing amounts of data an
organization needs to deal with today. Parallel databases are often quite expensive—
on the order of tens of thousands of dollars per terabyte of user data. Over the past
few years, Hadoop has gained popularity as a platform for data-warehousing. Ham-
merbacher [68], for example, discussed Facebook’s experiences with scaling up business
intelligence applications with Oracle databases, which they ultimately abandoned in
favor of a Hadoop-based solution developed in-house called Hive (which is now an
open-source project). Pig [114] is a platform for massive data analytics built on Hadoop
and capable of handling structured as well as semi-structured data. It was originally
developed by Yahoo, but is now also an open-source project.
Given successful applications of Hadoop to data-warehousing and complex ana-
lytical queries that are prevalent in such an environment, it makes sense to examine
MapReduce algorithms for manipulating relational data. This section focuses specif-
ically on performing relational joins in MapReduce. We should stress here that even
though Hadoop has been applied to process relational data, Hadoop is not a database.
There is an ongoing debate between advocates of parallel databases and proponents
of MapReduce regarding the merits of both approaches for OLAP-type workloads. De-
witt and Stonebraker, two well-known figures in the database community, famously
decried MapReduce as “a major step backwards” in a controversial blog post.11 With
colleagues, they ran a series of benchmarks that demonstrated the supposed superiority
of column-oriented parallel databases over Hadoop [120, 144]. However, see Dean and
Ghemawat’s counterarguments [47] and recent attempts at hybrid architectures [1].
We shall refrain here from participating in this lively debate, and instead focus on
discussing algorithms. From an application point of view, it is highly unlikely that an
analyst interacting with a data warehouse will ever be called upon to write MapReduce
programs (and indeed, Hadoop-based systems such as Hive and Pig present a much
higher-level language for interacting with large amounts of data). Nevertheless, it is
instructive to understand the algorithms that underlie basic relational operations.
This section presents three different strategies for performing relational joins on
two datasets (relations), generically named S and T . Let us suppose that relation S
looks something like the following:
(k1 , s1 , S1 )
(k2 , s2 , S2 )
(k3 , s3 , S3 )
...
where k is the key we would like to join on, sn is a unique id for the tuple, and the
Sn after sn denotes other attributes in the tuple (unimportant for the purposes of the
join). Similarly, suppose relation T looks something like this:
11 http://databasecolumn.vertica.com/database-innovation/mapreduce-a-major-step-backwards/
64 CHAPTER 3. MAPREDUCE ALGORITHM DESIGN
(k1 , t1 , T1 )
(k3 , t2 , T2 )
(k8 , t3 , T3 )
...
where k is the join key, tn is a unique id for the tuple, and the Tn after tn denotes other
attributes in the tuple.
To make this task more concrete, we present one realistic scenario: S might rep-
resent a collection of user profiles, in which case k could be interpreted as the primary
key (i.e., user id). The tuples might contain demographic information such as age, gen-
der, income, etc. The other dataset, T , might represent logs of online activity. Each
tuple might correspond to a page view of a particular URL and may contain additional
information such as time spent on the page, ad revenue generated, etc. The k in these
tuples could be interpreted as the foreign key that associates each individual page view
with a user. Joining these two datasets would allow an analyst, for example, to break
down online activity in terms of demographics.
basic MapReduce programming model, no guarantees are made about value ordering,
so the first value might be from S or from T . We can proceed to join the two tuples
and perform additional computations (e.g., filter by some other attribute, compute
aggregates, etc.). If there is only one value associated with a key, this means that no
tuple in the other dataset shares the join key, so the reducer does nothing.
Let us now consider the one-to-many join. Assume that tuples in S have unique
join keys (i.e., k is the primary key in S), so that S is the “one” and T is the “many”.
The above algorithm will still work, but when processing each key in the reducer, we
have no idea when the value corresponding to the tuple from S will be encountered, since
values are arbitrarily ordered. The easiest solution is to buffer all values in memory,
pick out the tuple from S, and then cross it with every tuple from T to perform the
join. However, as we have seen several times already, this creates a scalability bottleneck
since we may not have sufficient memory to hold all the tuples with the same join key.
This is a problem that requires a secondary sort, and the solution lies in the
value-to-key conversion design pattern we just presented. In the mapper, instead of
simply emitting the join key as the intermediate key, we instead create a composite key
consisting of the join key and the tuple id (from either S or T ). Two additional changes
are required: First, we must define the sort order of the keys to first sort by the join
key, and then sort all tuple ids from S before all tuple ids from T . Second, we must
define the partitioner to pay attention to only the join key, so that all composite keys
with the same join key arrive at the same reducer.
After applying the value-to-key conversion design pattern, the reducer will be
presented with keys and values along the lines of the following:
Since both the join key and the tuple id are present in the intermediate key, we can
remove them from the value to save a bit of space.13 Whenever the reducer encounters
a new join key, it is guaranteed that the associated value will be the relevant tuple from
S. The reducer can hold this tuple in memory and then proceed to cross it with tuples
from T in subsequent steps (until a new join key is encountered). Since the MapReduce
execution framework performs the sorting, there is no need to buffer tuples (other than
the single one from S). Thus, we have eliminated the scalability bottleneck.
Finally, let us consider the many-to-many join case. Assuming that S is the smaller
dataset, the above algorithm works as well. Consider what happens at the reducer:
13 Once again, not very important if the intermediate data is compressed.
66 CHAPTER 3. MAPREDUCE ALGORITHM DESIGN
partitioner and controlling the sort order of key-value pairs). For ad hoc data analysis,
reduce-side joins are a more general, albeit less efficient, solution. Consider the case
where datasets have multiple keys that one might wish to join on—then no matter
how the data is organized, map-side joins will require repartitioning of the data. Al-
ternatively, it is always possible to repartition a dataset using an identity mapper and
reducer. But of course, this incurs the cost of shuffling data over the network.
There is a final restriction to bear in mind when using map-side joins with the
Hadoop implementation of MapReduce. We assume here that the datasets to be joined
were produced by previous MapReduce jobs, so this restriction applies to keys the
reducers in those jobs may emit. Hadoop permits reducers to emit keys that are different
from the input key whose values they are processing (that is, input and output keys
need not be the same, nor even the same type).15 However, if the output key of a
reducer is different from the input key, then the output dataset from the reducer will
not necessarily be partitioned in a manner consistent with the specified partitioner
(because the partitioner applies to the input keys rather than the output keys). Since
map-side joins depend on consistent partitioning and sorting of keys, the reducers used
to generate data that will participate in a later map-side join must not emit any key
but the one they are currently processing.
15 Incontrast, recall from Section 2.2 that in Google’s implementation, reducers’ output keys must be exactly
same as their input keys.
68 CHAPTER 3. MAPREDUCE ALGORITHM DESIGN
3.6 SUMMARY
This chapter provides a guide on the design of MapReduce algorithms. In particular,
we present a number of “design patterns” that capture effective solutions to common
problems. In summary, they are:
• “In-mapper combining”, where the functionality of the combiner is moved into the
mapper. Instead of emitting intermediate output for every input key-value pair,
the mapper aggregates partial results across multiple input records and only emits
intermediate key-value pairs after some amount of local aggregation is performed.
• The related patterns “pairs” and “stripes” for keeping track of joint events from
a large number of observations. In the pairs approach, we keep track of each joint
event separately, whereas in the stripes approach we keep track of all events that
co-occur with the same event. Although the stripes approach is significantly more
efficient, it requires memory on the order of the size of the event space, which
presents a scalability bottleneck.
• “Order inversion”, where the main idea is to convert the sequencing of compu-
tations into a sorting problem. Through careful orchestration, we can send the
reducer the result of a computation (e.g., an aggregate statistic) before it encoun-
ters the data necessary to produce that computation.
• “Value-to-key conversion”, which provides a scalable solution for secondary sort-
ing. By moving part of the value into the key, we can exploit the MapReduce
execution framework itself for sorting.
Ultimately, controlling synchronization in the MapReduce programming model boils
down to effective use of the following techniques:
1. Constructing complex keys and values that bring together data necessary for a
computation. This is used in all of the above design patterns.
16 In order to achieve good performance in accessing distributed key-value stores, it is often necessary to batch
queries before making synchronous requests (to amortize latency over many requests) or to rely on asynchronous
requests.
3.6. SUMMARY 69
CHAPTER 4
1 Leavingaside the problem of searching live data streams such a tweets, which requires different techniques and
algorithms.
4.1. WEB CRAWLING 71
other hand, is an online problem that demands sub-second response time. Individual
users expect low query latencies, but query throughput is equally important since a
retrieval engine must usually serve many users concurrently. Furthermore, query loads
are highly variable, depending on the time of day, and can exhibit “spikey” behavior
due to special circumstances (e.g., a breaking news event triggers a large number of
searches on the same topic). On the other hand, resource consumption for the indexing
problem is more predictable.
A comprehensive treatment of web search is beyond the scope of this chapter,
and even this entire book. Explicitly recognizing this, we mostly focus on the problem
of inverted indexing, the task most amenable to solutions in MapReduce. This chapter
begins by first providing an overview of web crawling (Section 4.1) and introducing the
basic structure of an inverted index (Section 4.2). A baseline inverted indexing algorithm
in MapReduce is presented in Section 4.3. We point out a scalability bottleneck in that
algorithm, which leads to a revised version presented in Section 4.4. Index compression
is discussed in Section 4.5, which fills in missing details on building compact index
structures. Since MapReduce is primarily designed for batch-oriented processing, it
does not provide an adequate solution for the retrieval problem, an issue we discuss in
Section 4.6. The chapter concludes with a summary and pointers to additional readings.
further processing, and repeats. In fact, rudimentary web crawlers can be written in a
few hundred lines of code.
However, effective and efficient web crawling is far more complex. The following
lists a number of issues that real-world crawlers must contend with:
• A web crawler must practice good “etiquette” and not overload web servers. For
example, it is common practice to wait a fixed amount of time before repeated
requests to the same server. In order to respect these constraints while maintaining
good throughput, a crawler typically keeps many execution threads running in
parallel and maintains many TCP connections (perhaps hundreds) open at the
same time.
• Since a crawler has finite bandwidth and resources, it must prioritize the order in
which unvisited pages are downloaded. Such decisions must be made online and
in an adversarial environment, in the sense that spammers actively create “link
farms” and “spider traps” full of spam pages to trick a crawler into overrepresent-
ing content from a particular site.
• Most real-world web crawlers are distributed systems that run on clusters of ma-
chines, often geographically distributed. To avoid downloading a page multiple
times and to ensure data consistency, the crawler as a whole needs mechanisms
for coordination and load-balancing. It also needs to be robust with respect to
machine failures, network outages, and errors of various types.
• Web content changes, but with different frequency depending on both the site and
the nature of the content. A web crawler needs to learn these update patterns
to ensure that content is reasonably current. Getting the right recrawl frequency
is tricky: too frequent means wasted resources, but not frequent enough leads to
stale content.
• The web is full of duplicate content. Examples include multiple copies of a popu-
lar conference paper, mirrors of frequently-accessed sites such as Wikipedia, and
newswire content that is often duplicated. The problem is compounded by the fact
that most repetitious pages are not exact duplicates but near duplicates (that is,
basically the same page but with different ads, navigation bars, etc.) It is desir-
able during the crawling process to identify near duplicates and select the best
exemplar to index.
• The web is multilingual. There is no guarantee that pages in one language only
link to pages in the same language. For example, a professor in Asia may maintain
her website in the local language, but contain links to publications in English.
4.2. INVERTED INDEXES 73
Furthermore, many pages contain a mix of text in different languages. Since doc-
ument processing techniques (e.g., tokenization, stemming) differ by language, it
is important to identify the (dominant) language on a page.
The above discussion is not meant to be an exhaustive enumeration of issues, but rather
to give the reader an appreciation of the complexities involved in this intuitively simple
task. For more information, see a recent survey on web crawling [113]. Section 4.7
provides pointers to additional readings.
term1 d1 p d5 p d6 p d11 p …
terms postings
term1 d1 p d5 p d6 p d11 p …
Figure 4.1: Simple illustration of an inverted index. Each term is associated with a list of
postings. Each posting is comprised of a document id and a payload, denoted by p in this case.
An inverted index provides quick access to documents ids that contain a term.
the front of a postings list. Either way, an auxiliary data structure is necessary to
maintain the mapping from integer document ids to some other more meaningful handle,
such as a URL.
Given a query, retrieval involves fetching postings lists associated with query terms
and traversing the postings to compute the result set. In the simplest case, boolean
retrieval involves set operations (union for boolean OR and intersection for boolean
AND) on postings lists, which can be accomplished very efficiently since the postings
are sorted by document id. In the general case, however, query–document scores must be
computed. Partial document scores are stored in structures called accumulators. At the
end (i.e., once all postings have been processed), the top k documents are then extracted
to yield a ranked list of results for the user. Of course, there are many optimization
strategies for query evaluation (both approximate and exact) that reduce the number
of postings a retrieval engine must examine.
The size of an inverted index varies, depending on the payload stored in each
posting. If only term frequency is stored, a well-optimized inverted index can be a tenth
of the size of the original document collection. An inverted index that stores positional
information would easily be several times larger than one that does not. Generally, it
is possible to hold the entire vocabulary (i.e., dictionary of all the terms) in memory,
especially with techniques such as front-coding [156]. However, with the exception of
well-resourced, commercial web search engines,6 postings lists are usually too large to
store in memory and must be held on disk, usually in compressed form (more details in
Section 4.5). Query evaluation, therefore, necessarily involves random disk access and
“decoding” of the postings. One important aspect of the retrieval problem is to organize
disk operations such that random seeks are minimized.
Figure 4.2: Pseudo-code of the baseline inverted indexing algorithm in MapReduce. Map-
pers emit postings keyed by terms, the execution framework groups postings by term, and the
reducers write postings lists to disk.
Once again, this brief discussion glosses over many complexities and does a huge
injustice to the tremendous amount of research in information retrieval. However, our
goal is to provide the reader with an overview of the important issues; Section 4.7
provides references to additional readings.
After this histogram has been built, the mapper then iterates over all terms. For each
term, a pair consisting of the document id and the term frequency is created. Each pair,
denoted by hn, H{t}i in the pseudo-code, represents an individual posting. The mapper
then emits an intermediate key-value pair with the term as the key and the posting as
the value, in line 7 of the mapper pseudo-code. Although as presented here only the
term frequency is stored in the posting, this algorithm can be easily augmented to store
additional information (e.g., term positions) in the payload.
In the shuffle and sort phase, the MapReduce runtime essentially performs a large,
distributed group by of the postings by term. Without any additional effort by the
programmer, the execution framework brings together all the postings that belong in
the same postings list. This tremendously simplifies the task of the reducer, which
simply needs to gather together all the postings and write them to disk. The reducer
begins by initializing an empty list and then appends all postings associated with the
same key (term) to the list. The postings are then sorted by document id, and the entire
postings list is emitted as a value, with the term as the key. Typically, the postings list
is first compressed, but we leave this aside for now (see Section 4.4 for more details).
The final key-value pairs are written to disk and comprise the inverted index. Since
each reducer writes its output in a separate file in the distributed file system, our final
index will be split across r files, where r is the number of reducers. There is no need to
further consolidate these files. Separately, we must also build an index to the postings
lists themselves for the retrieval engine: this is typically in the form of mappings from
term to (file, byte offset) pairs, so that given a term, the retrieval engine can fetch
its postings list by opening the appropriate file and seeking to the correct byte offset
position in that file.
Execution of the complete algorithm is illustrated in Figure 4.3 with a toy example
consisting of three documents, three mappers, and two reducers. Intermediate key-value
pairs (from the mappers) and the final key-value pairs comprising the inverted index
(from the reducers) are shown in the boxes with dotted lines. Postings are shown as
pairs of boxes, with the document id on the left and the term frequency on the right.
The MapReduce programming model provides a very concise expression of the in-
verted indexing algorithm. Its implementation is similarly concise: the basic algorithm
can be implemented in as few as a couple dozen lines of code in Hadoop (with mini-
mal document processing). Such an implementation can be completed as a week-long
programming assignment in a course for advanced undergraduates or first-year gradu-
ate students [83, 93]. In a non-MapReduce indexer, a significant fraction of the code
is devoted to grouping postings by term, given constraints imposed by memory and
disk (e.g., memory capacity is limited, disk seeks are slow, etc.). In MapReduce, the
programmer does not need to worry about any of these issues—most of the heavy lifting
is performed by the execution framework.
4.4. INVERTED INDEXING: REVISED IMPLEMENTATION 77
reducer reducer
fish d1 2 d2 2 bird d3 1
one d1 1 d3 1 blue d2 1
two d1 1 red d2 1 d3 1
Figure 4.3: Simple illustration of the baseline inverted indexing algorithm in MapReduce with
three mappers and two reducers. Postings are shown as pairs of boxes (docid, tf).
7 See
similar discussion in Section 3.4: in principle, this need not be an in-memory sort. It is entirely possible to
implement a disk-based sort within the reducer.
78 CHAPTER 4. INVERTED INDEXING FOR TEXT RETRIEVAL
bottleneck is to let the MapReduce runtime do the sorting for us. Instead of emitting
key-value pairs of the following type:
In other words, the key is a tuple containing the term and the document id, while the
value is the term frequency. This is exactly the value-to-key conversion design pattern
introduced in Section 3.4. With this modification, the programming model ensures that
the postings arrive in the correct order. This, combined with the fact that reducers can
hold state across multiple keys, allows postings lists to be created with minimal memory
usage. As a detail, remember that we must define a custom partitioner to ensure that
all tuples with the same term are shuffled to the same reducer.
The revised MapReduce inverted indexing algorithm is shown in Figure 4.4. The
mapper remains unchanged for the most part, other than differences in the intermediate
key-value pairs. The Reduce method is called for each key (i.e., ht, ni), and by design,
there will only be one value associated with each key. For each key-value pair, a posting
can be directly added to the postings list. Since the postings are guaranteed to arrive
in sorted order by document id, they can be incrementally coded in compressed form—
thus ensuring a small memory footprint. Finally, when all postings associated with the
same term have been processed (i.e., t 6= tprev ), the entire postings list is emitted. The
final postings list must be written out in the Close method. As with the baseline
algorithm, payloads can be easily changed: by simply replacing the intermediate value
f (term frequency) with whatever else is desired (e.g., term positional information).
There is one more detail we must address when building inverted indexes. Since
almost all retrieval models take into account document length when computing query–
document scores, this information must also be extracted. Although it is straightforward
to express this computation as another MapReduce job, this task can actually be folded
into the inverted indexing process. When processing the terms in each document, the
document length is known, and can be written out as “side data” directly to HDFS.
We can take advantage of the ability for a mapper to hold state across the processing of
multiple documents in the following manner: an in-memory associative array is created
to store document lengths, which is populated as each document is processed.8 When
the mapper finishes processing input records, document lengths are written out to
HDFS (i.e., in the Close method). This approach is essentially a variant of the in-
mapper combining pattern. Document length data ends up in m different files, where
m is the number of mappers; these files are then consolidated into a more compact
8 In general, there is no worry about insufficient memory to hold these data.
4.5. INDEX COMPRESSION 79
1: class Mapper
2: method Map(docid n, doc d)
3: H ← new AssociativeArray
4: for all term t ∈ doc d do
5: H{t} ← H{t} + 1
6: for all term t ∈ H do
7: Emit(tuple ht, ni, tf H{t})
1: class Reducer
2: method Initialize
3: tprev ← ∅
4: P ← new PostingsList
5: method Reduce(tuple ht, ni, tf [f ])
6: if t 6= tprev ∧ tprev 6= ∅ then
7: Emit(term t, postings P )
8: P.Reset()
9: P.Add(hn, f i)
10: tprev ← t
11: method Close
12: Emit(term t, postings P )
integer9 and the second as a 16-bit integer. Thus, a postings list might be encoded as
follows:
[(5, 2), (7, 3), (12, 1), (49, 1), (51, 2), . . .]
where each posting is represented by a pair in parentheses. Note that all brackets, paren-
theses, and commas are only included to enhance readability; in reality the postings
would be represented as a long stream of integers. This naı̈ve implementation would
require six bytes per posting. Using this scheme, the entire inverted index would be
about as large as the collection itself. Fortunately, we can do significantly better.
The first trick is to encode differences between document ids as opposed to the
document ids themselves. Since the postings are sorted by document ids, the differences
(called d-gaps) must be positive integers greater than zero. The above postings list,
represented with d-gaps, would be:
[(5, 2), (2, 3), (5, 1), (37, 1), (2, 2), . . .]
Of course, we must actually encode the first document id. We haven’t lost any infor-
mation, since the original document ids can be easily reconstructed from the d-gaps.
However, it’s not obvious that we’ve reduced the space requirements either, since the
largest possible d-gap is one less than the number of documents in the collection.
This is where the second trick comes in, which is to represent the d-gaps in a
way such that it takes less space for smaller numbers. Similarly, we want to apply the
same techniques to compress the term frequencies, since for the most part they are also
small values. But to understand how this is done, we need to take a slight detour into
compression techniques, particularly for coding integers.
Compression, in general, can be characterized as either lossless or lossy: it’s fairly
obvious that loseless compression is required in this context. To start, it is important
to understand that all compression techniques represent a time–space tradeoff. That
is, we reduce the amount of space on disk necessary to store data, but at the cost of
extra processor cycles that must be spent coding and decoding data. Therefore, it is
possible that compression reduces size but also slows processing. However, if the two
factors are properly balanced (i.e., decoding speed can keep up with disk bandwidth),
we can achieve the best of both worlds: smaller and faster.
both occupy four bytes. Obviously, encoding d-gaps this way doesn’t yield any reduc-
tions in size.
A simple approach to compression is to only use as many bytes as is necessary to
represent the integer. This is known as variable-length integer coding (varInt for short)
and accomplished by using the high order bit of every byte as the continuation bit,
which is set to one in the last byte and zero elsewhere. As a result, we have 7 bits per
byte for coding the value, which means that 0 ≤ n < 27 can be expressed with 1 byte,
27 ≤ n < 214 with 2 bytes, 214 ≤ n < 221 with 3, and 221 ≤ n < 228 with 4 bytes. This
scheme can be extended to code arbitrarily-large integers (i.e., beyond 4 bytes). As a
concrete example, the two numbers:
127, 128
would be coded as such:
1 1111111, 0 0000001 1 0000000
The above code contains two code words, the first consisting of 1 byte, and the second
consisting of 2 bytes. Of course, the comma and the spaces are there only for readability.
Variable-length integers are byte-aligned because the code words always fall along byte
boundaries. As a result, there is never any ambiguity about where one code word ends
and the next begins. However, the downside of varInt coding is that decoding involves
lots of bit operations (masks, shifts). Furthermore, the continuation bit sometimes re-
sults in frequent branch mispredicts (depending on the actual distribution of d-gaps),
which slows down processing.
A variant of the varInt scheme was described by Jeff Dean in a keynote talk at
the WSDM 2009 conference.10 The insight is to code groups of four integers at a time.
Each group begins with a prefix byte, divided into four 2-bit values that specify the
byte length of each of the following integers. For example, the following prefix byte:
00,00,01,10
indicates that the following four integers are one byte, one byte, two bytes, and three
bytes, respectively. Therefore, each group of four integers would consume anywhere be-
tween 5 and 17 bytes. A simple lookup table based on the prefix byte directs the decoder
on how to process subsequent bytes to recover the coded integers. The advantage of this
group varInt coding scheme is that values can be decoded with fewer branch mispredicts
and bitwise operations. Experiments reported by Dean suggest that decoding integers
with this scheme is more than twice as fast as the basic varInt scheme.
In most architectures, accessing entire machine words is more efficient than fetch-
ing all its bytes separately. Therefore, it makes sense to store postings in increments
10 http://research.google.com/people/jeff/WSDM09-keynote.pdf
82 CHAPTER 4. INVERTED INDEXING FOR TEXT RETRIEVAL
of 16-bit, 32-bit, or 64-bit machine words. Anh and Moffat [8] presented several word-
aligned coding methods, one of which is called Simple-9, based on 32-bit words. In this
coding scheme, four bits in each 32-bit word are reserved as a selector. The remaining
28 bits are used to code actual integer values. Now, there are a variety of ways these 28
bits can be divided to code one or more integers: 28 bits can be used to code one 28-bit
integer, two 14-bit integers, three 9-bit integers (with one bit unused), etc., all the way
up to twenty-eight 1-bit integers. In fact, there are nine different ways the 28 bits can be
divided into equal parts (hence the name of the technique), some with leftover unused
bits. This is stored in the selector bits. Therefore, decoding involves reading a 32-bit
word, examining the selector to see how the remaining 28 bits are packed, and then
appropriately decoding each integer. Coding works in the opposite way: the algorithm
scans ahead to see how many integers can be squeezed into 28 bits, packs those integers,
and sets the selector bits appropriately.
0001101001010100
00 01 1 01 00 1 01 01 00
Golomb
x unary γ b=5 b = 10
1 0 0 0:00 0:000
2 10 10:0 0:01 0:001
3 110 10:1 0:10 0:010
4 1110 110:00 0:110 0:011
5 11110 110:01 0:111 0:100
6 111110 110:10 10:00 0:101
7 1111110 110:11 10:01 0:1100
8 11111110 1110:000 10:10 0:1101
9 111111110 1110:001 10:110 0:1110
10 1111111110 1110:010 10:111 0:1111
Figure 4.5: The first ten positive integers in unary, γ, and Golomb (b = 5, 10) codes.
of zero, which is fine since d-gaps and term frequencies should never be zero.11 As an
example, 4 in unary code is 1110. With unary code we can code x in x bits, which
although economical for small values, becomes inefficient for even moderately large
values. Unary codes are rarely used by themselves, but form a component of other
coding schemes. Unary codes of the first ten positive integers are shown in Figure 4.5.
Elias γ code is an efficient coding scheme that is widely used in practice. An integer
x > 0 is broken into two components, 1 + blog2 xc (= n, the length), which is coded in
unary code, and x − 2blog2 xc (= r, the remainder), which is in binary.12 The unary
component n specifies the number of bits required to code x, and the binary component
codes the remainder r in n − 1 bits. As an example, consider x = 10: 1 + blog2 10c =
4, which is 1110. The binary component codes x − 23 = 2 in 4 − 1 = 3 bits, which is
010. Putting both together, we arrive at 1110:010. The extra colon is inserted only for
readability; it’s not part of the final code, of course.
Working in reverse, it is easy to unambiguously decode a bit stream of γ codes:
First, we read a unary code cu , which is a prefix code. This tells us that the binary
portion is written in cu − 1 bits, which we then read as cb . We can then reconstruct x
as 2cu −1 + cb . For x < 16, γ codes occupy less than a full byte, which makes them more
compact than variable-length integer codes. Since term frequencies for the most part are
relatively small, γ codes make sense for them and can yield substantial space savings.
For reference, the γ codes of the first ten positive integers are shown in Figure 4.5. A
11 As a note, some sources describe slightly different formulations of the same coding scheme. Here, we adopt the
conventions used in the classic IR text Managing Gigabytes [156].
12 Note that bxc is the floor function, which maps x to the largest integer not greater than x, so, e.g., b3.8c = 3.
This is the default behavior in many programming languages when casting from a floating-point type to an
integer type.
84 CHAPTER 4. INVERTED INDEXING FOR TEXT RETRIEVAL
variation on γ code is δ code, where the n portion of the γ code is coded in γ code
itself (as opposed to unary code). For smaller values γ codes are more compact, but for
larger values, δ codes take less space.
Unary and γ codes are parameterless, but even better compression can be achieved
with parameterized codes. A good example of this is Golomb code. For some parameter
b, an integer x > 0 is coded in two parts: first, we compute q = b(x − 1)/bc and code
q + 1 in unary; then, we code the remainder r = x − qb − 1 in truncated binary. This
is accomplished as follows: if b is a power of two, then truncated binary is exactly the
same as normal binary, requiring log2 b bits. Otherwise, we code the first 2blog2 bc+1 − b
values of r in blog2 bc bits and code the rest of the values of r by coding r + 2blog2 bc+1 − b
in ordinary binary representation using blog2 bc + 1 bits. In this case, the r is coded in
either blog2 bc or blog2 bc + 1 bits, and unlike ordinary binary coding, truncated binary
codes are prefix codes. As an example, if b = 5, then r can take the values {0, 1, 2, 3, 4},
which would be coded with the following code words: {00, 01, 10, 110, 111}. For reference,
Golomb codes of the first ten positive integers are shown in Figure 4.5 for b = 5 and
b = 10. A special case of Golomb code is worth noting: if b is a power of two, then
coding and decoding can be handled more efficiently (needing only bit shifts and bit
masks, as opposed to multiplication and division). These are known as Rice codes.
Researchers have shown that Golomb compression works well for d-gaps, and is
optimal with the following parameter setting:
df
b ≈ 0.69 × (4.1)
N
where df is the document frequency of the term, and N is the number of documents in
the collection.13
Putting everything together, one popular approach for postings compression is to
represent d-gaps with Golomb codes and term frequencies with γ codes [156, 162]. If
positional information is desired, we can use the same trick to code differences between
term positions using γ codes.
the subset of documents processed by the mapper). Once the mapper has processed all
input records, special keys of the form ht, ∗i are emitted with the partial df as the value.
To ensure that these special keys arrive first, we define the sort order of the
tuple so that the special symbol ∗ precedes all documents (part of the order inversion
design pattern). Thus, for each term, the reducer will first encounter the ht, ∗i key,
associated with a list of values representing partial df values originating from each
mapper. Summing all these partial contributions will yield the term’s df, which can
then be used to set the Golomb compression parameter b. This allows the postings to be
incrementally compressed as they are encountered in the reducer—memory bottlenecks
are eliminated since we do not need to buffer postings in memory.
Once again, the order inversion design pattern comes to the rescue. Recall that
the pattern is useful when a reducer needs to access the result of a computation (e.g.,
an aggregate statistic) before it encounters the data necessary to produce that compu-
tation. For computing relative frequencies, that bit of information was the marginal. In
this case, it’s the document frequency.
d1 d2 d3 d4 d5 d6 d7 d8 d9
t1 2 3
t2 1 1 4 partitiona
t3 1 1 2
t4 5 2 2
t5 1 1 3 partitionb
t6 2 1
t7 2 1 4
t8 1 2 3 partitionc
t9 1 2 1
Figure 4.6: Term–document matrix for a toy collection (nine documents, nine terms) illus-
trating different partitioning strategies: partitioning vertically (1, 2, 3) corresponds to document
partitioning, whereas partitioning horizontally (a, b, c) corresponds to term partitioning.
server holds the complete index for a subset of the entire collection. This corresponds
to partitioning vertically in Figure 4.6. With term partitioning, on the other hand,
each server is responsible for a subset of the terms for the entire collection. That is, a
server holds the postings for all documents in the collection for a subset of terms. This
corresponds to partitioning horizontally in Figure 4.6.
Document and term partitioning require different retrieval strategies and represent
different tradeoffs. Retrieval under document partitioning involves a query broker, which
forwards the user’s query to all partition servers, merges partial results from each, and
then returns the final results to the user. With this architecture, searching the entire
collection requires that the query be processed by every partition server. However,
since each partition operates independently and traverses postings in parallel, document
partitioning typically yields shorter query latencies (compared to a single monolithic
index with much longer postings lists).
Retrieval under term partitioning, on the other hand, requires a very different
strategy. Suppose the user’s query Q contains three terms, q1 , q2 , and q3 . Under the
88 CHAPTER 4. INVERTED INDEXING FOR TEXT RETRIEVAL
pipelined query evaluation strategy, the broker begins by forwarding the query to the
server that holds the postings for q1 (usually the least frequent term). The server tra-
verses the appropriate postings list and computes partial query–document scores, stored
in the accumulators. The accumulators are then passed to the server that holds the post-
ings associated with q2 for additional processing, and then to the server for q3 , before
final results are passed back to the broker and returned to the user. Although this
query evaluation strategy may not substantially reduce the latency of any particular
query, it can theoretically increase a system’s throughput due to the far smaller number
of total disk seeks required for each user query (compared to document partitioning).
However, load-balancing is tricky in a pipelined term-partitioned architecture due to
skew in the distribution of query terms, which can create “hot spots” on servers that
hold the postings for frequently-occurring query terms.
In general, studies have shown that document partitioning is a better strategy
overall [109], and this is the strategy adopted by Google [16]. Furthermore, it is known
that Google maintains its indexes in memory (although this is certainly not the common
case for search engines in general). One key advantage of document partitioning is
that result quality degrades gracefully with machine failures. Partition servers that are
offline will simply fail to deliver results for their subsets of the collection. With sufficient
partitions, users might not even be aware that documents are missing. For most queries,
the web contains more relevant documents than any user has time to digest: users of
course care about getting relevant documents (sometimes, they are happy with a single
relevant document), but they are generally less discriminating when it comes to which
relevant documents appear in their results (out of the set of all relevant documents).
Note that partitions may be unavailable due to reasons other than machine failure:
cycling through different partitions is a very simple and non-disruptive strategy for
index updates.
Working in a document-partitioned architecture, there are a variety of approaches
to dividing up the web into smaller pieces. Proper partitioning of the collection can
address one major weakness of this architecture, which is that every partition server
is involved in every user query. Along one dimension, it is desirable to partition by
document quality using one or more classifiers; see [124] for a recent survey on web
page classification. Partitioning by document quality supports a multi-phase search
strategy: the system examines partitions containing high quality documents first, and
only backs off to partitions containing lower quality documents if necessary. This reduces
the number of servers that need to be contacted for a user query. Along an orthogonal
dimension, it is desirable to partition documents by content (perhaps also guided by
the distribution of user queries from logs), so that each partition is “well separated”
from the others in terms of topical coverage. This also reduces the number of machines
that need to be involved in serving a user’s query: the broker can direct queries only to
4.7. SUMMARY AND ADDITIONAL READINGS 89
the partitions that are likely to contain relevant documents, as opposed to forwarding
the user query to all the partitions.
On a large-scale, reliability of service is provided by replication, both in terms
of multiple machines serving the same partition within a single datacenter, but also
replication across geographically-distributed datacenters. This creates at least two query
routing problems: since it makes sense to serve clients from the closest datacenter, a
service must route queries to the appropriate location. Within a single datacenter, the
system needs to properly balance load across replicas.
There are two final components of real-world search engines that are worth dis-
cussing. First, recall that postings only store document ids. Therefore, raw retrieval
results consist of a ranked list of semantically meaningless document ids. It is typically
the responsibility of document servers, functionally distinct from the partition servers
holding the indexes, to generate meaningful output for user presentation. Abstractly, a
document server takes as input a query and a document id, and computes an appropri-
ate result entry, typically comprising the title and URL of the page, a snippet of the
source document showing the user’s query terms in context, and additional metadata
about the document. Second, query evaluation can benefit immensely from caching, of
individual postings (assuming that the index is not already in memory) and even results
of entire queries [13]. This is made possible by the Zipfian distribution of queries, with
very frequent queries at the head of the distribution dominating the total number of
queries. Search engines take advantage of this with cache servers, which are functionally
distinct from all of the components discussed above.
compressing d-gaps and γ codes for term frequencies. We showed how the order inver-
sion design pattern introduced in Section 3.3 for computing relative frequencies can be
used to properly set compression parameters.
Additional Readings. Our brief discussion of web search glosses over many com-
plexities and does a huge injustice to the tremendous amount of research in information
retrieval. Here, however, we provide a few entry points into the literature. A survey ar-
ticle by Zobel and Moffat [162] is an excellent starting point on indexing and retrieval
algorithms. Another by Baeza-Yates et al. [11] overviews many important issues in
distributed retrieval. A keynote talk at the WSDM 2009 conference by Jeff Dean re-
vealed a lot of information about the evolution of the Google search architecture.15
Finally, a number of general information retrieval textbooks have been recently pub-
lished [101, 42, 30]. Of these three, the one by Büttcher et al. [30] is noteworthy in
having detailed experimental evaluations that compare the performance (both effec-
tiveness and efficiency) of a wide range of algorithms and techniques. While outdated
in many other respects, the textbook Managing Gigabytes [156] remains an excellent
source for index compression techniques. Finally, ACM SIGIR is an annual conference
and the most prestigious venue for academic information retrieval research; proceedings
from those events are perhaps the best starting point for those wishing to keep abreast
of publicly-documented developments in the field.
15 http://research.google.com/people/jeff/WSDM09-keynote.pdf
91
CHAPTER 5
Graph Algorithms
Graphs are ubiquitous in modern society: examples encountered by almost everyone
on a daily basis include the hyperlink structure of the web (simply known as the web
graph), social networks (manifest in the flow of email, phone call patterns, connections
on social networking sites, etc.), and transportation networks (roads, bus routes, flights,
etc.). Our very own existence is dependent on an intricate metabolic and regulatory
network, which can be characterized as a large, complex graph involving interactions
between genes, proteins, and other cellular products. This chapter focuses on graph
algorithms in MapReduce. Although most of the content has nothing to do with text
processing per se, documents frequently exist in the context of some underlying network,
making graph analysis an important component of many text processing applications.
Perhaps the best known example is PageRank, a measure of web page quality based
on the structure of hyperlinks, which is used in ranking results for web search. As one
of the first applications of MapReduce, PageRank exemplifies a large class of graph
algorithms that can be concisely captured in the programming model. We will discuss
PageRank in detail later this chapter.
In general, graphs can be characterized by nodes (or vertices) and links (or edges)
that connect pairs of nodes.1 These connections can be directed or undirected. In some
graphs, there may be an edge from a node to itself, resulting in a self loop; in others,
such edges are disallowed. We assume that both nodes and links may be annotated with
additional metadata: as a simple example, in a social network where nodes represent
individuals, there might be demographic information (e.g., age, gender, location) at-
tached to the nodes and type information attached to the links (e.g., indicating type of
relationship such as “friend” or “spouse”).
Mathematicians have always been fascinated with graphs, dating back to Euler’s
paper on the Seven Bridges of Königsberg in 1736. Over the past few centuries, graphs
have been extensively studied, and today much is known about their properties. Far
more than theoretical curiosities, theorems and algorithms on graphs can be applied to
solve many real-world problems:
• Graph search and path planning. Search algorithms on graphs are invoked millions
of times a day, whenever anyone searches for directions on the web. Similar algo-
rithms are also involved in friend recommendations and expert-finding in social
networks. Path planning problems involving everything from network packets to
delivery trucks represent another large class of graph search problems.
1 Throughout this chapter, we use node interchangeably with vertex and similarly with link and edge.
92 CHAPTER 5. GRAPH ALGORITHMS
• Graph clustering. Can a large graph be divided into components that are relatively
disjoint (for example, as measured by inter-component links [59])? Among other
applications, this task is useful for identifying communities in social networks (of
interest to sociologists who wish to understand how human relationships form and
evolve) and for partitioning large graphs (of interest to computer scientists who
seek to better parallelize graph processing). See [158] for a survey.
• Minimum spanning trees. A minimum spanning tree for a graph G with weighted
edges is a tree that contains all vertices of the graph and a subset of edges that
minimizes the sum of edge weights. A real-world example of this problem is a
telecommunications company that wishes to lay optical fiber to span a number
of destinations at the lowest possible cost (where weights denote costs). This ap-
proach has also been applied to wide variety of problems, including social networks
and the migration of Polynesian islanders [64].
• Bipartite graph matching. A bipartite graph is one whose vertices can be divided
into two disjoint sets. Matching problems on such graphs can be used to model
job seekers looking for employment or singles looking for dates.
• Maximum flow. In a weighted directed graph with two special nodes called the
source and the sink, the max flow problem involves computing the amount of
“traffic” that can be sent from source to sink given various flow capacities defined
by edge weights. Transportation companies (airlines, shipping, etc.) and network
operators grapple with complex versions of these problems on a daily basis.
• Identifying “special” nodes. There are many ways to define what special means,
including metrics based on node in-degree, average distance to other nodes, and
relationship to cluster structure. These special nodes are important to investigators
attempting to break up terrorist cells, epidemiologists modeling the spread of
diseases, advertisers trying to promote products, and many others.
A common feature of these problems is the scale of the datasets on which the algorithms
must operate: for example, the hyperlink structure of the web, which contains billions
of pages, or social networks that contain hundreds of millions of individuals. Clearly,
algorithms that run on a single machine and depend on the entire graph residing in
memory are not scalable. We’d like to put MapReduce to work on these challenges.2
This chapter is organized as follows: we begin in Section 5.1 with an introduction
to graph representations, and then explore two classic graph algorithms in MapReduce:
2 As a side note, Google recently published a short description of a system called Pregel [98], based on Valiant’s
Bulk Synchronous Parallel model [148], for large-scale graph algorithms; a longer description is anticipated in
a forthcoming paper [99]
5.1. GRAPH REPRESENTATIONS 93
n2
n1 n2 n3 n4 n5
n1
n1 0 1 0 1 0 n1 [n2, n4]
n2 0 0 1 0 1 n2 [n3, n5]
n3 0 0 0 1 0 n3 [n4]
n5 n4 0 0 0 0 1 n4 [n5]
n3
n5 1 1 1 0 0 n5 [n1, n2, n3]
Figure 5.1: A simple directed graph (left) represented as an adjacency matrix (middle) and
with adjacency lists (right).
parallel breadth-first search (Section 5.2) and PageRank (Section 5.3). Before conclud-
ing with a summary and pointing out additional readings, Section 5.4 discusses a number
of general issue that affect graph processing with MapReduce.
web. In this chapter, we assume processing of sparse graphs, although we will return to
this issue in Section 5.4.
The major problem with an adjacency matrix representation for sparse graphs
is its O(n2 ) space requirement. Furthermore, most of the cells are zero, by definition.
As a result, most computational implementations of graph algorithms operate over
adjacency lists, in which a node is associated with neighbors that can be reached via
outgoing edges. Figure 5.1 also shows the adjacency list representation of the graph
under consideration (on the right). For example, since n1 is connected by directed
edges to n2 and n4 , those two nodes will be on the adjacency list of n1 . There are two
options for encoding undirected graphs: one could simply encode each edge twice (if ni
and nj are connected, each appears on each other’s adjacency list). Alternatively, one
could order the nodes (arbitrarily or otherwise) and encode edges only on the adjacency
list of the node that comes first in the ordering (i.e., if i < j, then nj is on the adjacency
list of ni , but not the other way around).
Note that certain graph operations are easier on adjacency matrices than on ad-
jacency lists. In the first, operations on incoming links for each node translate into a
column scan on the matrix, whereas operations on outgoing links for each node trans-
late into a row scan. With adjacency lists, it is natural to operate on outgoing links, but
computing anything that requires knowledge of the incoming links of a node is difficult.
However, as we shall see, the shuffle and sort mechanism in MapReduce provides an
easy way to group edges by their destination nodes, thus allowing us to compute over
incoming edges with in the reducer. This property of the execution framework can also
be used to invert the edges of a directed graph, by mapping over the nodes’ adjacency
lists and emitting key–value pairs with the destination node id as the key and the source
node id as the value.4
4 Thistechnique is used in anchor text inversion, where one gathers the anchor text of hyperlinks pointing to a
particular page. It is common practice to enrich a web page’s standard textual representation with all of the
anchor text associated with its incoming hyperlinks (e.g., [107]).
5.2. PARALLEL BREADTH-FIRST SEARCH 95
1: Dijkstra(G, w, s)
2: d[s] ← 0
3: for all vertex v ∈ V do
4: d[v] ← ∞
5: Q ← {V }
6: while Q 6= ∅ do
7: u ← ExtractMin(Q)
8: for all vertex v ∈ u.AdjacencyList do
9: if d[v] > d[u] + w(u, v) then
10: d[v] ← d[u] + w(u, v)
Figure 5.2: Pseudo-code for Dijkstra’s algorithm, which is based on maintaining a global
priority queue of nodes with priorities equal to their distances from the source node. At each
iteration, the algorithm expands the node with the shortest distance and updates distances to
all reachable nodes.
n2 n4 n2 n4 n2 n4
∞ 1
∞ 10 1
∞ 8 1
14
10 10 10
9 9 9
2 3 2 3 2 3
0 4 6
0 4 6
0 4 6
n1 n1 n1
5 7 5 7 5 7
∞ ∞ 5 ∞ 5 7
2 2 2
n3 n5 n3 n5 n3 n5
(a) (b) (c)
n2 n4 n2 n4 n2 n4
1 1 1
8 13 8 9 8 9
10 10 10
9 9 9
2 3 2 3 2 3
0 4 6
0 4 6
0 4 6
n1 n1 n1
5 7 5 7 5 7
5 7 5 7 5 7
2 2 2
n3 n5 n3 n5 n3 n5
(d) (e) (f)
Figure 5.3: Example of Dijkstra’s algorithm applied to a simple graph with five nodes, with n1
as the source and edge distances as indicated. Parts (a)–(e) show the running of the algorithm
at each iteration, with the current distance inside the node. Nodes with thicker borders are
those being expanded; nodes that have already been expanded are shown in black.
(c) that the distance to n2 has decreased because we’ve found a shorter path. The nodes
that will be expanded next, in order, are n5 , n2 , and n4 . The algorithm terminates with
the end state shown in (f), where we’ve discovered the shortest distance to all nodes.
The key to Dijkstra’s algorithm is the priority queue that maintains a globally-
sorted list of nodes by current distance. This is not possible in MapReduce, as the
programming model does not provide a mechanism for exchanging global data. Instead,
we adopt a brute force approach known as parallel breadth-first search. First, as a
simplification let us assume that all edges have unit distance (modeling, for example,
hyperlinks on the web). This makes the algorithm easier to understand, but we’ll relax
this restriction later.
The intuition behind the algorithm is this: the distance of all nodes connected
directly to the source node is one; the distance of all nodes directly connected to those
is two; and so on. Imagine water rippling away from a rock dropped into a pond—
that’s a good image of how parallel breadth-first search works. However, what if there
are multiple paths to the same node? Suppose we wish to compute the shortest distance
5.2. PARALLEL BREADTH-FIRST SEARCH 97
to node n. The shortest path must go through one of the nodes in M that contains an
outgoing edge to n: we need to examine all m ∈ M to find ms , the node with the shortest
distance. The shortest distance to n is the distance to ms plus one.
Pseudo-code for the implementation of the parallel breadth-first search algorithm
is provided in Figure 5.4. As with Dijkstra’s algorithm, we assume a connected, directed
graph represented as adjacency lists. Distance to each node is directly stored alongside
the adjacency list of that node, and initialized to ∞ for all nodes except for the source
node. In the pseudo-code, we use n to denote the node id (an integer) and N to denote
the node’s corresponding data structure (adjacency list and current distance). The
algorithm works by mapping over all nodes and emitting a key-value pair for each
neighbor on the node’s adjacency list. The key contains the node id of the neighbor,
and the value is the current distance to the node plus one. This says: if we can reach node
n with a distance d, then we must be able to reach all the nodes that are connected to
n with distance d + 1. After shuffle and sort, reducers will receive keys corresponding to
the destination node ids and distances corresponding to all paths leading to that node.
The reducer will select the shortest of these distances and then update the distance in
the node data structure.
It is apparent that parallel breadth-first search is an iterative algorithm, where
each iteration corresponds to a MapReduce job. The first time we run the algorithm, we
“discover” all nodes that are connected to the source. The second iteration, we discover
all nodes connected to those, and so on. Each iteration of the algorithm expands the
“search frontier” by one hop, and, eventually, all nodes will be discovered with their
shortest distances (assuming a fully-connected graph). Before we discuss termination
of the algorithm, there is one more detail required to make the parallel breadth-first
search algorithm work. We need to “pass along” the graph structure from one iteration
to the next. This is accomplished by emitting the node data structure itself, with the
node id as a key (Figure 5.4, line 4 in the mapper). In the reducer, we must distinguish
the node data structure from distance values (Figure 5.4, lines 5–6 in the reducer), and
update the minimum distance in the node data structure before emitting it as the final
value. The final output is now ready to serve as input to the next iteration.5
So how many iterations are necessary to compute the shortest distance to all
nodes? The answer is the diameter of the graph, or the greatest distance between any
pair of nodes. This number is surprisingly small for many real-world problems: the
saying “six degrees of separation” suggests that everyone on the planet is connected to
everyone else by at most six steps (the people a person knows are one step away, people
that they know are two steps away, etc.). If this is indeed true, then parallel breadth-
first search on the global social network would take at most six MapReduce iterations.
5 Notethat in this algorithm we are overloading the value type, which can either be a distance (integer) or a
complex data structure representing a node. The best way to achieve this in Hadoop is to create a wrapper
object with an indicator variable specifying what the content is.
98 CHAPTER 5. GRAPH ALGORITHMS
1: class Mapper
2: method Map(nid n, node N )
3: d ← N.Distance
4: Emit(nid n, N ) . Pass along graph structure
5: for all nodeid m ∈ N.AdjacencyList do
6: Emit(nid m, d + 1) . Emit distances to reachable nodes
1: class Reducer
2: method Reduce(nid m, [d1 , d2 , . . .])
3: dmin ← ∞
4: M ←∅
5: for all d ∈ counts [d1 , d2 , . . .] do
6: if IsNode(d) then
7: M ←d . Recover graph structure
8: else if d < dmin then . Look for shorter distance
9: dmin ← d
10: M.Distance ← dmin . Update shortest distance
11: Emit(nid m, node M )
Figure 5.4: Pseudo-code for parallel breath-first search in MapReduce: the mappers emit dis-
tances to reachable nodes, while the reducers select the minimum of those distances for each
destination node. Each iteration (one MapReduce job) of the algorithm expands the “search
frontier” by one hop.
For more serious academic studies of “small world” phenomena in networks, we refer
the reader to a number of publications [61, 62, 152, 2]. In practical terms, we iterate
the algorithm until there are no more node distances that are ∞. Since the graph is
connected, all nodes are reachable, and since all edge distances are one, all discovered
nodes are guaranteed to have the shortest distances (i.e., there is not a shorter path
that goes through a node that hasn’t been discovered).
The actual checking of the termination condition must occur outside of Map-
Reduce. Typically, execution of an iterative MapReduce algorithm requires a non-
MapReduce “driver” program, which submits a MapReduce job to iterate the algorithm,
checks to see if a termination condition has been met, and if not, repeats. Hadoop pro-
vides a lightweight API for constructs called “counters”, which, as the name suggests,
can be used for counting events that occur during execution, e.g., number of corrupt
records, number of times a certain condition is met, or anything that the programmer
desires. Counters can be defined to count the number of nodes that have distances of
∞: at the end of the job, the driver program can access the final counter value and
check to see if another iteration is necessary.
5.2. PARALLEL BREADTH-FIRST SEARCH 99
search frontier
s
q
p
Figure 5.5: In the single source shortest path problem with arbitrary edge distances, the
shortest path from source s to node r may go outside the current search frontier, in which case
we will not find the shortest distance to r until the search frontier expands to cover q.
Finally, as with Dijkstra’s algorithm in the form presented earlier, the parallel
breadth-first search algorithm only finds the shortest distances, not the actual shortest
paths. However, the path can be straightforwardly recovered. Storing “backpointers”
at each node, as with Dijkstra’s algorithm, will work, but may not be efficient since
the graph needs to be traversed again to reconstruct the path segments. A simpler
approach is to emit paths along with distances in the mapper, so that each node will
have its shortest path easily accessible at all times. The additional space requirements
for shuffling these data from mappers to reducers are relatively modest, since for the
most part paths (i.e., sequence of node ids) are relatively short.
Up until now, we have been assuming that all edges are unit distance. Let us relax
that restriction and see what changes are required in the parallel breadth-first search
algorithm. The adjacency lists, which were previously lists of node ids, must now encode
the edge distances as well. In line 6 of the mapper code in Figure 5.4, instead of emitting
d + 1 as the value, we must now emit d + w where w is the edge distance. No other
changes to the algorithm are required, but the termination behavior is very different.
To illustrate, consider the graph fragment in Figure 5.5, where s is the source node,
and in this iteration, we just “discovered” node r for the very first time. Assume for
the sake of argument that we’ve already discovered the shortest distance to node p, and
that the shortest distance to r so far goes through p. This, however, does not guarantee
that we’ve discovered the shortest distance to r, since there may exist a path going
through q that we haven’t encountered yet (because it lies outside the search frontier).6
However, as the search frontier expands, we’ll eventually cover q and all other nodes
along the path from p to q to r—which means that with sufficient iterations, we will
discover the shortest distance to r. But how do we know that we’ve found the shortest
distance to p? Well, if the shortest path to p lies within the search frontier, we would
6 Note that the same argument does not apply to the unit edge distance case: the shortest path cannot lie outside
the search frontier since any such path would necessarily be longer.
100 CHAPTER 5. GRAPH ALGORITHMS
1
1 1
n6 n7
n8
10
1
n5 n9
n1
1 1
1 n4
n2 1
n3
Figure 5.6: A sample graph that elicits worst-case behavior for parallel breadth-first search.
Eight iterations are required to discover shortest distances to all nodes from n1 .
have already discovered it. And if it doesn’t, the above argument applies. Similarly, we
can repeat the same argument for all nodes on the path from s to p. The conclusion is
that, with sufficient iterations, we’ll eventually discover all the shortest distances.
So exactly how many iterations does “eventually” mean? In the worst case, we
might need as many iterations as there are nodes in the graph minus one. In fact, it
is not difficult to construct graphs that will elicit this worse-case behavior: Figure 5.6
provides an example, with n1 as the source. The parallel breadth-first search algorithm
would not discover that the shortest path from n1 to n6 goes through n3 , n4 , and n5
until the fifth iteration. Three more iterations are necessary to cover the rest of the
graph. Fortunately, for most real-world graphs, such extreme cases are rare, and the
number of iterations necessary to discover all shortest distances is quite close to the
diameter of the graph, as in the unit edge distance case.
In practical terms, how do we know when to stop iterating in the case of arbitrary
edge distances? The algorithm can terminate when shortest distances at every node no
longer change. Once again, we can use counters to keep track of such events. Every time
we encounter a shorter distance in the reducer, we increment a counter. At the end of
each MapReduce iteration, the driver program reads the counter value and determines
if another iteration is necessary.
Compared to Dijkstra’s algorithm on a single processor, parallel breadth-first
search in MapReduce can be characterized as a brute force approach that “wastes” a
lot of time performing computations whose results are discarded. At each iteration, the
algorithm attempts to recompute distances to all nodes, but in reality only useful work
is done along the search frontier: inside the search frontier, the algorithm is simply
repeating previous computations.7 Outside the search frontier, the algorithm hasn’t
7 Unlessthe algorithm discovers an instance of the situation described in Figure 5.5, in which case, updated
distances will propagate inside the search frontier.
5.2. PARALLEL BREADTH-FIRST SEARCH 101
discovered any paths to nodes there yet, so no meaningful work is done. Dijkstra’s
algorithm, on the other hand, is far more efficient. Every time a node is explored, we’re
guaranteed to have already found the shortest path to it. However, this is made possible
by maintaining a global data structure (a priority queue) that holds nodes sorted by
distance—this is not possible in MapReduce because the programming model does not
provide support for global data that is mutable and accessible by the mappers and
reducers. These inefficiencies represent the cost of parallelization.
The parallel breadth-first search algorithm is instructive in that it represents the
prototypical structure of a large class of graph algorithms in MapReduce. They share
in the following characteristics:
• The graph structure is represented with adjacency lists, which is part of some larger
node data structure that may contain additional information (variables to store
intermediate output, features of the nodes). In many cases, features are attached
to edges as well (e.g., edge weights).
• The MapReduce algorithm maps over the node data structures and performs a
computation that is a function of features of the node, intermediate output at-
tached to each node, and features of the adjacency list (outgoing edges and their
features). In other words, computations can only involve a node’s internal state
and its local graph structure. The results of these computations are emitted as val-
ues, keyed with the node ids of the neighbors (i.e., those nodes on the adjacency
lists). Conceptually, we can think of this as “passing” the results of the computa-
tion along outgoing edges. In the reducer, the algorithm receives all partial results
that have the same destination node, and performs another computation (usually,
some form of aggregation).
• In addition to computations, the graph itself is also passed from the mapper to the
reducer. In the reducer, the data structure corresponding to each node is updated
and written back to disk.
• Graph algorithms in MapReduce are generally iterative, where the output of the
previous iteration serves as input to the next iteration. The process is controlled
by a non-MapReduce driver program that checks for termination.
For parallel breadth-first search, the mapper computation is the current distance plus
edge distance (emitting distances to neighbors), while the reducer computation is the
Min function (selecting the shortest path). As we will see in the next section, the
MapReduce algorithm for PageRank works in much the same way.
102 CHAPTER 5. GRAPH ALGORITHMS
5.3 PAGERANK
PageRank [117] is a measure of web page quality based on the structure of the hyperlink
graph. Although it is only one of thousands of features that is taken into account in
Google’s search algorithm, it is perhaps one of the best known and most studied.
A vivid way to illustrate PageRank is to imagine a random web surfer: the surfer
visits a page, randomly clicks a link on that page, and repeats ad infinitum. Page-
Rank is a measure of how frequently a page would be encountered by our tireless web
surfer. More precisely, PageRank is a probability distribution over nodes in the graph
representing the likelihood that a random walk over the link structure will arrive at a
particular node. Nodes that have high in-degrees tend to have high PageRank values,
as well as nodes that are linked to by other nodes with high PageRank values. This
behavior makes intuitive sense: if PageRank is a measure of page quality, we would ex-
pect high-quality pages to contain “endorsements” from many other pages in the form
of hyperlinks. Similarly, if a high-quality page links to another page, then the second
page is likely to be high quality also. PageRank represents one particular approach to
inferring the quality of a web page based on hyperlink structure; two other popular
algorithms, not covered here, are SALSA [88] and HITS [84] (also known as “hubs and
authorities”).
The complete formulation of PageRank includes an additional component. As it
turns out, our web surfer doesn’t just randomly click links. Before the surfer decides
where to go next, a biased coin is flipped—heads, the surfer clicks on a random link on
the page as usual. Tails, however, the surfer ignores the links on the page and randomly
“jumps” or “teleports” to a completely different page.
But enough about random web surfing. Formally, the PageRank P of a page n is
defined as follows:
1 X P (m)
P (n) = α + (1 − α) (5.1)
|G| m∈L(n)
C(m)
where |G| is the total number of nodes (pages) in the graph, α is the random jump
factor, L(n) is the set of pages that link to n, and C(m) is the out-degree of node m
(the number of links on page m). The random jump factor α is sometimes called the
“teleportation” factor; alternatively, (1 − α) is referred to as the “damping” factor.
Let us break down each component of the formula in detail. First, note that
PageRank is defined recursively—this gives rise to an iterative algorithm we will detail
in a bit. A web page n receives PageRank “contributions” from all pages that link to
it, L(n). Let us consider a page m from the set of pages L(n): a random surfer at
m will arrive at n with probability 1/C(m) since a link is selected at random from all
outgoing links. Since the PageRank value of m is the probability that the random surfer
will be at m, the probability of arriving at n from m is P (m)/C(m). To compute the
5.3. PAGERANK 103
PageRank of n, we need to sum contributions from all pages that link to n. This is
the summation in the second half of the equation. However, we also need to take into
account the random jump: there is a 1/|G| chance of landing at any particular page,
where |G| is the number of nodes in the graph. Of course, the two contributions need to
be combined: with probability α the random surfer executes a random jump, and with
probability 1 − α the random surfer follows a hyperlink.
Note that PageRank assumes a community of honest users who are not trying to
“game” the measure. This is, of course, not true in the real world, where an adversarial
relationship exists between search engine companies and a host of other organizations
and individuals (marketers, spammers, activists, etc.) who are trying to manipulate
search results—to promote a cause, product, or service, or in some cases, to trap and
intentionally deceive users (see, for example, [12, 63]). A simple example is a so-called
“spider trap”, a infinite chain of pages (e.g., generated by CGI) that all link to a single
page (thereby artificially inflating its PageRank). For this reason, PageRank is only one
of thousands of features used in ranking web pages.
The fact that PageRank is recursively defined translates into an iterative algo-
rithm which is quite similar in basic structure to parallel breadth-first search. We start
by presenting an informal sketch. At the beginning of each iteration, a node passes its
PageRank contributions to other nodes that it is connected to. Since PageRank is a
probability distribution, we can think of this as spreading probability mass to neigh-
bors via outgoing links. To conclude the iteration, each node sums up all PageRank
contributions that have been passed to it and computes an updated PageRank score.
We can think of this as gathering probability mass passed to a node via its incoming
links. This algorithm iterates until PageRank values don’t change anymore.
Figure 5.7 shows a toy example that illustrates two iterations of the algorithm.
As a simplification, we ignore the random jump factor for now (i.e., α = 0) and further
assume that there are no dangling nodes (i.e., nodes with no outgoing edges). The
algorithm begins by initializing a uniform distribution of PageRank values across nodes.
In the beginning of the first iteration (top, left), partial PageRank contributions are sent
from each node to its neighbors connected via outgoing links. For example, n1 sends
0.1 PageRank mass to n2 and 0.1 PageRank mass to n4 . This makes sense in terms of
the random surfer model: if the surfer is at n1 with a probability of 0.2, then the surfer
could end up either in n2 or n4 with a probability of 0.1 each. The same occurs for all
the other nodes in the graph: note that n5 must split its PageRank mass three ways,
since it has three neighbors, and n4 receives all the mass belonging to n3 because n3
isn’t connected to any other node. The end of the first iteration is shown in the top
right: each node sums up PageRank contributions from its neighbors. Note that since
n1 has only one incoming link, from n3 , its updated PageRank value is smaller than
before, i.e., it “passed along” more PageRank mass than it received. The exact same
104 CHAPTER 5. GRAPH ALGORITHMS
It
Iteration
ti 1 n2 (0.2) n2 (0.166)
0.1
n1 (0.2) 0.1 0.1 n1 (0.066)
0.1
0.066
0.066 0.066
n5 (0.2) n5 (0.3)
n3 (0.2) n3 (0.166)
0.2 0.2
n4 (0.2) n4 (0.3)
0.033
0.1
01
0.1 01
0.1
n5 (0.3) n5 (0.383)
n3 (0.166) n3 (0.183)
0.3 0.166
n4 (0.3)
(0 3) n4 (0.2)
(0 2)
Figure 5.7: PageRank toy example showing two iterations, top and bottom. Left graphs show
PageRank values at the beginning of each iteration and how much PageRank mass is passed to
each neighbor. Right graphs show updated PageRank values at the end of each iteration.
process repeats, and the second iteration in our toy example is illustrated by the bottom
two graphs. At the beginning of each iteration, the PageRank values of all nodes sum
to one. PageRank mass is preserved by the algorithm, guaranteeing that we continue
to have a valid probability distribution at the end of each iteration.
Pseudo-code of the MapReduce PageRank algorithm is shown in Figure 5.8; it is
simplified in that we continue to ignore the random jump factor and assume no dangling
nodes (complications that we will return to later). An illustration of the running algo-
rithm is shown in Figure 5.9 for the first iteration of the toy graph in Figure 5.7. The
algorithm maps over the nodes, and for each node computes how much PageRank mass
needs to be distributed to its neighbors (i.e., nodes on the adjacency list). Each piece
of the PageRank mass is emitted as the value, keyed by the node ids of the neighbors.
Conceptually, we can think of this as passing PageRank mass along outgoing edges.
In the shuffle and sort phase, the MapReduce execution framework groups values
(piece of PageRank mass) passed along the graph edges by destination node (i.e., all
edges that point to the same node). In the reducer, PageRank mass contributions from
all incoming edges are summed to arrive at the updated PageRank value for each node.
5.3. PAGERANK 105
1: class Mapper
2: method Map(nid n, node N )
3: p ← N.PageRank/|N.AdjacencyList|
4: Emit(nid n, N ) . Pass along graph structure
5: for all nodeid m ∈ N.AdjacencyList do
6: Emit(nid m, p) . Pass PageRank mass to neighbors
1: class Reducer
2: method Reduce(nid m, [p1 , p2 , . . .])
3: M ←∅
4: for all p ∈ counts [p1 , p2 , . . .] do
5: if IsNode(p) then
6: M ←p . Recover graph structure
7: else
8: s←s+p . Sum incoming PageRank contributions
9: M.PageRank ← s
10: Emit(nid m, node M )
Figure 5.8: Pseudo-code for PageRank in MapReduce (leaving aside dangling nodes and the
random jump factor). In the map phase we evenly divide up each node’s PageRank mass and
pass each piece along outgoing edges to neighbors. In the reduce phase PageRank contributions
are summed up at each destination node. Each MapReduce job corresponds to one iteration of
the algorithm.
Iteration 1 n2 (0.2) n2 (0.166)
0.1
n1 ((0.2)) 0.1 0.1 n1 ((0.066))
0.1
0.066
0.066 0.066
n5 (0.2) n5 (0.3)
n3 (0.2)
(0 2) n3 (0.166)
(0 166)
0.2 0.2
n4 (0.2)
106 CHAPTER 5. GRAPH ALGORITHMS n4 (0.3)
Map
n2 n4 n3 n5 n4 n5 n1 n2 n3
n1 n2 n2 n3 n3 n4 n4 n5 n5
Reduce
Figure 5.9: Illustration of the MapReduce PageRank algorithm corresponding to the first
iteration in Figure 5.7. The size of each box is proportion to its PageRank value. During the
map phase, PageRank mass is distributed evenly to nodes on each node’s adjacency list (shown
at the very top). Intermediate values are keyed by node (shown inside the boxes). In the reduce
phase, all partial PageRank contributions are summed together to arrive at updated values.
As with the parallel breadth-first search algorithm, the graph structure itself must be
passed from iteration to iteration. Each node data structure is emitted in the mapper
and written back out to disk in the reducer. All PageRank mass emitted by the mappers
are accounted for in the reducer: since we begin with the sum of PageRank values across
all nodes equal to one, the sum of all the updated PageRank values should remain a
valid probability distribution.
Having discussed the simplified PageRank algorithm in MapReduce, let us now
take into account the random jump factor and dangling nodes: as it turns out both are
treated similarly. Dangling nodes are nodes in the graph that have no outgoing edges,
i.e., their adjacency lists are empty. In the hyperlink graph of the web, these might
correspond to pages in a crawl that have not been downloaded yet. If we simply run
the algorithm in Figure 5.8 on graphs with dangling nodes, the total PageRank mass
will not be conserved, since no key-value pairs will be emitted when a dangling node is
encountered in the mappers.
The proper treatment of PageRank mass “lost” at the dangling nodes is to re-
distribute it across all nodes in the graph evenly (cf. [22]). There are many ways to
determine the missing PageRank mass. One simple approach is by instrumenting the
algorithm in Figure 5.8 with counters: whenever the mapper processes a node with an
empty adjacency list, it keeps track of the node’s PageRank value in the counter. At
the end of the iteration, we can access the counter to find out how much PageRank
5.3. PAGERANK 107
mass was lost at the dangling nodes. Another approach is to reserve a special key for
8
storing PageRank mass from dangling nodes. When the mapper encounters a dangling
node, its PageRank mass is emitted with the special key; the reducer must be modified
to contain special logic for handling the missing PageRank mass. Yet another approach
is to write out the missing PageRank mass as “side data” for each map task (using
the in-mapper combining technique for aggregation); a final pass in the driver program
is needed to sum the mass across all map tasks. Either way, we arrive at the amount
of PageRank mass lost at the dangling nodes—this then must be redistribute evenly
across all nodes.
This redistribution process can be accomplished by mapping over all nodes again.
At the same time, we can take into account the random jump factor. For each node, its
current PageRank value p is updated to the final PageRank value p0 according to the
following formula:
1 m
p =α
0
+ (1 − α) +p (5.2)
|G| |G|
where m is the missing PageRank mass, and |G| is the number of nodes in the entire
graph. We add the PageRank mass from link traversal (p, computed from before) to
the share of the lost PageRank mass that is distributed to each node (m/|G|). Finally,
we take into account the random jump factor: with probability α the random surfer
arrives via jumping, and with probability 1 − α the random surfer arrives via incoming
links. Note that this MapReduce job requires no reducers.
Putting everything together, one iteration of PageRank requires two MapReduce
jobs: the first to distribute PageRank mass along graph edges, and the second to take
care of dangling nodes and the random jump factor. At end of each iteration, we end
up with exactly the same data structure as the beginning, which is a requirement for
the iterative algorithm to work. Also, the PageRank values of all nodes sum up to one,
which ensures a valid probability distribution.
Typically, PageRank is iterated until convergence, i.e., when the PageRank values
of nodes no longer change (within some tolerance, to take into account, for example,
floating point precision errors). Therefore, at the end of each iteration, the PageRank
driver program must check to see if convergence has been reached. Alternative stopping
criteria include running a fixed number of iterations (useful if one wishes to bound
algorithm running time) or stopping when the ranks of PageRank values no longer
change. The latter is useful for some applications that only care about comparing the
PageRank of two arbitrary pages and do not need the actual PageRank values. Rank
stability is obtained faster than the actual convergence of values.
8 InHadoop, counters are 8-byte integers: a simple workaround is to multiply PageRank values by a large
constant, and then cast as an integer.
108 CHAPTER 5. GRAPH ALGORITHMS
In absolute terms, how many iterations are necessary for PageRank to converge?
This is a difficult question to precisely answer since it depends on many factors, but
generally, fewer than one might expect. In the original PageRank paper [117], conver-
gence on a graph with 322 million edges was reached in 52 iterations (see also Bianchini
et al. [22] for additional discussion). On today’s web, the answer is not very meaningful
due to the adversarial nature of web search as previously discussed—the web is full
of spam and populated with sites that are actively trying to “game” PageRank and
related hyperlink-based metrics. As a result, running PageRank in its unmodified form
presented here would yield unexpected and undesirable results. Of course, strategies
developed by web search companies to combat link spam are proprietary (and closely-
guarded secrets, for obvious reasons)—but undoubtedly these algorithmic modifications
impact convergence behavior. A full discussion of the escalating “arms race” between
search engine companies and those that seek to promote their sites is beyond the scope
of this book.9
if one wishes to propagate information from a node to all nodes that are within two links, one could process
graph G to derive graph G0 , where there would exist a link from node ni to nj if nj was reachable within two
link traversals of ni in the original graph G.
5.4. ISSUES WITH GRAPH PROCESSING 109
(
b + log(1 + ea−b ) a < b
a⊕b=
a + log(1 + eb−a ) a ≥ b
Furthermore, many math libraries include a log1p function which computes log(1 + x)
with higher precision than the naı̈ve implementation would have when x is very small
(as is often the case when working with probabilities). Its use may further improve the
accuracy of implementations that use log probabilities.
alignment and assembly with graph algorithms in MapReduce. Finally, it is easy to for-
get that parallel graph algorithms have been studied by computer scientists for several
decades, particular in the PRAM model [77, 60]. It is not clear, however, to what extent
well-known PRAM algorithms translate naturally into the MapReduce framework.
112
CHAPTER 6
engineered system. As a result, correcting errors that the trained system makes can be
quite challenging.
Data-driven information processing systems can be constructed using a variety
of mathematical techniques, but in this chapter we focus on statistical models, which
probabilistically relate inputs from an input set X (e.g., sentences, documents, etc.),
which are always observable, to annotations from a set Y, which is the space of possible
annotations or analyses that the system should predict. This model may take the form
of either a joint model Pr(x, y) which assigns a probability to every pair hx, yi ∈ X ×
Y or a conditional model Pr(y|x), which assigns a probability to every y ∈ Y, given
an x ∈ X . For example, to create a statistical spam detection system, we might have
Y = {Spam, NotSpam} and X be the set of all possible email messages. For machine
translation, X might be the set of Arabic sentences and Y the set of English sentences.1
There are three closely related, but distinct challenges in statistical text-
processing. The first is model selection. This entails selecting a representation of a
joint or conditional distribution over the desired X and Y. For a problem where X and
Y are very small, one could imagine representing these probabilities in look-up tables.
However, for something like email classification or machine translation, where the model
space is infinite, the probabilities cannot be represented directly, and must be computed
algorithmically. As an example of such models, we introduce hidden Markov models
(HMMs), which define a joint distribution over sequences of inputs and sequences of
annotations. The second challenge is parameter estimation or learning, which involves
the application of a optimization algorithm and training criterion to select the param-
eters of the model to optimize the model’s performance (with respect to the given
training criterion) on the training data.2 The parameters of a statistical model are
the values used to compute the probability of some event described by the model. In
this chapter we will focus on one particularly simple training criterion for parameter
estimation, maximum likelihood estimation, which says to select the parameters that
make the training data most probable under the model, and one learning algorithm
that attempts to meet this criterion, called expectation maximization (EM). The final
challenge for statistical modeling is the problem of decoding, or, given some x, using the
model to select an annotation y. One very common strategy is to select y according to
the following criterion:
y ∗ = arg max Pr(y|x)
y∈Y
1 In this chapter, we will consider discrete models only. They tend to be sufficient for text processing, and their
presentation is simpler than models with continuous densities. It should be kept in mind that the sets X and
Y may still be countably infinite.
2 We restrict our discussion in this chapter to models with finite numbers of parameters and where the learning
process refers to setting those parameters. Inference in and learning of so-called nonparameteric models, which
have an infinite number of parameters and have become important statistical models for text processing in
recent years, is beyond the scope of this chapter.
114 CHAPTER 6. EM ALGORITHMS FOR TEXT PROCESSING
In a conditional (or direct) model, this is a straightforward search for the best y un-
der the model. In a joint model, the search is also straightforward, on account of the
definition of conditional probability:
Pr(x, y)
y ∗ = arg max Pr(y|x) = arg max P = arg max Pr(x, y)
y 0 Pr(x, y )
y∈Y y∈Y 0 y∈Y
The specific form that the search takes will depend on how the model is represented.
Our focus in this chapter will primarily be on the second problem: learning parameters
for models, but we will touch on the third problem as well.
Machine learning is often categorized as either supervised or unsupervised. Su-
pervised learning of statistical models simply means that the model parameters are
estimated from training data consisting of pairs of inputs and annotations, that is
Z = hhx1 , y1 i, hx2 , y2 i, . . .i where hxi , yi i ∈ X × Y and yi is the gold standard (i.e., cor-
rect) annotation of xi . While supervised models often attain quite good performance,
they are often uneconomical to use, since the training data requires each object that
is to be classified (to pick a specific task), xi to be paired with its correct label, yi .
In many cases, these gold standard training labels must be generated by a process of
expert annotation, meaning that each xi must be manually labeled by a trained indi-
vidual. Even when the annotation task is quite simple for people to carry out (e.g., in
the case of spam detection), the number of potential examples that could be classified
(representing a subset of X , which may of course be infinite in size) will far exceed
the amount of data that can be annotated. As the annotation task becomes more com-
plicated (e.g., when predicting more complex structures such as sequences of labels or
when the annotation task requires specialized expertise), annotation becomes far more
challenging.
Unsupervised learning, on the other hand, requires only that the training data
consist of a representative collection of objects that should be annotated, that is
Z = hx1 , x2 , . . .i where xi ∈ X , but without any example annotations. While it may
at first seem counterintuitive that meaningful annotations can be learned without any
examples of the desired annotations being given, the learning criteria and model struc-
ture (which crucially define the space of possible annotations Y and the process by
which annotations relate to observable inputs) make it possible to induce annotations
by relying on regularities in the unclassified training instances. While a thorough discus-
sion of unsupervised learning is beyond the scope of this book, we focus on a particular
class of algorithms—expectation maximization (EM) algorithms—that can be used to
learn the parameters of a joint model Pr(x, y) from incomplete data (i.e., data where
some of the variables in the model cannot be observed; in the case of unsupervised
learning, the yi ’s are unobserved). Expectation maximization algorithms fit naturally
into the MapReduce paradigm, and are used to solve a number of problems of interest
in text processing. Furthermore, these algorithms can be quite computationally expen-
6.1. EXPECTATION MAXIMIZATION 115
sive, since they generally require repeated evaluations of the training data. MapReduce
therefore provides an opportunity not only to scale to larger amounts of data, but also
to improve efficiency bottlenecks at scales where non-parallel solutions could be utilized.
This chapter is organized as follows. In Section 6.1, we describe maximum likeli-
hood estimation for statistical models, show how this is generalized to models where not
all variables are observable, and then introduce expectation maximization (EM). We
describe hidden Markov models (HMMs) in Section 6.2, a very versatile class of models
that uses EM for parameter estimation. Section 6.3 discusses how EM algorithms can be
expressed in MapReduce, and then in Section 6.4 we look at a case study of word align-
ment for statistical machine translation. Section 6.5 examines similar algorithms that
are appropriate for supervised learning tasks. This chapter concludes with a summary
and pointers to additional readings.
To illustrate, consider the simple marble game shown in Figure 6.1. In this game,
a marble is released at the position indicated by the black dot, and it bounces down
into one of the cups at the bottom of the board, being diverted to the left or right by
the peg (indicated by a triangle) in the center. Our task is to construct a model that
predicts which cup the ball will drop into. A “rule-based” approach might be to take
116 CHAPTER 6. EM ALGORITHMS FOR TEXT PROCESSING
0 1
a b
0 1 2 3
a b c
Figure 6.1: A simple marble game where a released marble takes one of two possible paths.
This game can be modeled using a Bernoulli random variable with parameter p, which indicates
the probability that the marble will go to the right when it hits the peg.
exact measurements of the board and construct a physical model that we can use to
predict the behavior of the ball. Given sophisticated enough measurements, this could
certainly lead to a very accurate model. However, the construction of this model would
be quite time consuming and difficult.
A statistical approach, on the other hand, might be to assume that the behavior
of the marble in this game can be modeled using a Bernoulli random variable Y with
parameter p. That is, the value of the random variable indicates whether path 0 or 1 is
taken. We also define a random variable X whose value is the label of the cup that the
marble ends up in; note that X is deterministically related to Y , so an observation of
X is equivalent to an observation of Y .
To estimate the parameter p of the statistical model of our game, we need
some training data, so we drop 10 marbles into the game which end up in cups
x = hb, b, b, a, b, b, b, b, b, ai.
What is the maximum likelihood estimate of p given this data? By assuming
that our samples are independent and identically distributed (i.i.d.), we can write the
likelihood of our data as follows:3
10
Y
Pr(x; p) = pδ(xj ,a) (1 − p)δ(xj ,b)
j=1
= p2 · (1 − p)8
Since log is a monotonically increasing function, maximizing log Pr(x; p) will give us
the desired result. We can do this differentiating with respect to p and finding where
3 Inthis equation, δ is the Kroneker delta function which evaluates to 1 where its arguments are equal and 0
otherwise.
6.1. EXPECTATION MAXIMIZATION 117
d log Pr(x; p)
= 0
dp
d[2 · log p + 8 · log(1 − p)]
= 0
dp
2 8
− = 0
p 1−p
Solving for p yields 0.2, which is the intuitive result. Furthermore, it is straightforward
to show that in N trials where N0 marbles followed path 0 to cup a, and N1 marbles
followed path 1 to cup b, the maximum likelihood estimate of p is N1 /(N0 + N1 ).
While this model only makes use of an approximation of the true physical process
at work when the marble interacts with the game board, it is an empirical question
whether the model works well enough in practice to be useful. Additionally, while a
Bernoulli trial is an extreme approximation of the physical process, if insufficient re-
sources were invested in building a physical model, the approximation may perform
better than the more complicated “rule-based” model. This sort of dynamic is found
often in text processing problems: given enough data, astonishingly simple models can
outperform complex knowledge-intensive models that attempt to simulate complicated
processes.
0 1
a b
0 1 2 3
a b c
Figure 6.2: A more complicated marble game where the released marble takes one of four
possible paths. We assume that we can only observe which cup the marble ends up in, not the
specific path taken.
N2 + N3 N1 N3
p0 = p1 = p2 =
N N0 + N1 N2 + N3
However, we wish to consider the case where the paths taken are unobservable (imagine
an opaque sheet covering the center of the game board), but where we can see what cup a
marble ends in. In other words, we want to consider the case where we have partial data.
This is exactly the problem encountered in unsupervised learning: there is a statistical
model describing the relationship between two sets of variables (X’s and Y ’s), and there
is data available from just one of them. Furthermore, such algorithms are quite useful in
text processing, where latent variables may describe latent linguistic structures of the
observed variables, such as parse trees or part-of-speech tags, or alignment structures
relating sets of observed variables (see Section 6.4).
X
Pr(X = x) = Pr(X = x, Y = y; θ)
y∈Y
|x|
Y X
Pr(x; θ) = Pr(X = xj , Y = y; θ)
j=1 y∈Y
Thus, the maximum (marginal) likelihood estimate of the model parameters θ∗ given a
vector of i.i.d. observations x becomes:
|x|
Y X
θ = arg max
∗
Pr(X = xj , Y = y; θ)
θ
j=1 y∈Y
Unfortunately, in many cases, this maximum cannot be computed analytically, but the
iterative hill-climbing approach of expectation maximization can be used instead.
|x| |x|
Y X Y X
Pr(X = xj , Y = y; θ(i+1) ) ≥ Pr(X = xj , Y = y; θ(i) )
j=1 y∈Y j=1 y∈Y
The algorithm starts with some initial set of parameters θ(0) and then updates them
using two steps: expectation (E-step), which computes the posterior distribution over
the latent variables given the observable data x and a set of parameters θ(i) ,5 and
maximization (M-step), which computes new parameters θ(i+1) maximizing the expected
log likelihood of the joint distribution with respect to the distribution computed in the
E-step. The process then repeats with these new parameters. The algorithm terminates
when the likelihood remains unchanged.6 In more detail, the steps are as follows:
5 The term ‘expectation’ is used since the values computed in terms of the posterior distribution Pr(y|x; θ(i) )
that are required to solve the M-step have the form of an expectation (with respect to this distribution).
6 The final solution is only guaranteed to be a local maximum, but if the model is fully convex, it will also be
E-step. Compute the posterior probability of each possible hidden variable assign-
ments y ∈ Y for each x ∈ X and the current parameter settings, weighted by the rela-
tive frequency with which x occurs in x. Call this q(X = x, Y = y; θ(i) ) and note that
P
it defines a joint probability distribution over X × Y in that (x,y)∈X ×Y q(x, y) = 1.
Pr(x, y; θ(i) )
q(x, y; θ(i) ) = f (x|x) · Pr(Y = y|X = x; θ(i) ) = f (x|x) · P
y 0 Pr(x, y ; θ )
0 (i)
M-step. Compute new parameter settings that maximize the expected log of the
probability of the joint distribution under the q-distribution that was computed in the
E-step:
We omit the proof that the model with parameters θ(i+1) will have equal or greater
marginal likelihood on the training data than the model with parameters θ(i) , but this
is provably true [78].
Before continuing, we note that the effective application of expectation maximiza-
tion requires that both the E-step and the M-step consist of tractable computations.
Specifically, summing over the space of hidden variable assignments must not be in-
tractable. Depending on the independence assumptions made in the model, this may
be achieved through techniques such as dynamic programming. However, some models
may require intractable computations.
6.1.5 AN EM EXAMPLE
Let’s look at how to estimate the parameters from our latent variable marble game from
Section 6.1.2 using EM. We assume training data x consisting of N = |x| observations of
X with Na , Nb , and Nc indicating the number of marbles ending in cups a, b, and c. We
(0) (0) (0)
start with some parameters θ(0) = hp0 , p1 , p2 i that have been randomly initialized
to values between 0 and 1.
Nx
f (x|x) =
N
6.2. HIDDEN MARKOV MODELS 121
Next, we observe that Pr(Y = 0|X = a) = 1 and Pr(Y = 3|X = c) = 1 since cups a and
c fully determine the value of the path variable Y . The posterior probability of paths 1
and 2 are only non-zero when X is b:
data that are ordered sequentially (temporally, from left to right, etc.), such as words
in a sentence, base pairs in a gene, or letters in a word. These simple but powerful
models have been used in applications as diverse as speech recognition [78], information
extraction [139], gene finding [143], part of speech tagging [44], stock market forecasting
[70], text retrieval [108], and word alignment of parallel (translated) texts [150] (more
in Section 6.4).
In an HMM, the data being modeled is posited to have been generated from an
underlying Markov process, which is a stochastic process consisting of a finite set of
states where the probability of entering a state at time t + 1 depends only on the state
of the process at time t [130]. Alternatively, one can view a Markov process as a prob-
abilistic variant of a finite state machine, where transitions are taken probabilistically.
As another point of comparison, the PageRank algorithm considered in the previous
chapter (Section 5.3) can be understood as a Markov process: the probability of follow-
ing any link on a particular page is independent of the path taken to reach that page.
The states of this Markov process are, however, not directly observable (i.e., hidden).
Instead, at each time step, an observable token (e.g., a word, base pair, or letter) is
emitted according to a probability distribution conditioned on the identity of the state
that the underlying process is in.
A hidden Markov model M is defined as a tuple hS, O, θi. S is a finite set of states,
which generate symbols from a finite observation vocabulary O. Following convention,
we assume that variables q, r, and s refer to states in S, and o refers to symbols in
the observation vocabulary O. This model is parameterized by the tuple θ = hA, B, πi
consisting of an |S| × |S| matrix A of transition probabilities, where Aq (r) gives the
probability of transitioning from state q to state r; an |S| × |O| matrix B of emission
probabilities, where Bq (o) gives the probability that symbol o will be emitted from
state q; and an |S|-dimensional vector π, where πq is the probability that the process
starts in state q.7 These matrices may be dense, but for many applications sparse
parameterizations are useful. We further stipulate that Aq (r) ≥ 0, Bq (o) ≥ 0, and πq ≥ 0
for all q, r, and o, as well as that:
X X X
Aq (r) = 1 ∀q Bq (o) = 1 ∀q πq = 1
r∈S o∈O q∈S
Using our definition of an HMM, the answers to the first two questions are in principle
quite trivial to compute: by iterating over all state sequences Y, the probability that
8 The organization of this section is based in part on ideas from Lawrence Rabiner’s HMM tutorial [125].
124 CHAPTER 6. EM ALGORITHMS FOR TEXT PROCESSING
Initial probabilities:
DET ADJ NN V
0.5 0.1 0.3 0.1
Transition probabilities:
DET ADJ NN V
ADJ
DET 0 0 0 0.5
ADJ 0.3 0.2 0.1 0.2 NN
DET
NN 0.7 0.7 0.4 0.2
V 0 0.1 0.5 0.1
V
Emission probabilities:
DET ADJ NN V
Example outputs:
John might wash
NN V V
Figure 6.3: An example HMM that relates part-of-speech tags to vocabulary items in an
English-like language. Possible (probability > 0) transitions for the Markov process are shown
graphically. In the example outputs, the state sequences corresponding to the emissions are
written beneath the emitted symbols.
6.2. HIDDEN MARKOV MODELS 125
each generated x can be computed by looking up and multiplying the relevant prob-
abilities in A, B, and π, and then summing the result or taking the maximum. And,
as we hinted at in the previous section, the third question can be answered using EM.
Unfortunately, even with all the distributed computing power MapReduce makes avail-
able, we will quickly run into trouble if we try to use this naı̈ve strategy since there are
|S|τ distinct state sequences of length τ , making exhaustive enumeration computation-
ally intractable. Fortunately, because the underlying model behaves exactly the same
whenever it is in some state, regardless of how it got to that state, we can use dynamic
programming algorithms to answer all of the above questions without summing over
exponentially many sequences.
α1 (q) = πq · Bq (x1 )
From this, it’s not hard to see that the values of α2 (r) for every r can be computed in
terms of the |S| values in α1 (·) and the observation x2 :
X
α2 (r) = Br (x2 ) · α1 (q) · Aq (r)
q∈S
This works because there are |S| different ways to get to state r at time t = 2: starting
from state 1, 2, . . . , |S| and transitioning to state r. Furthermore, because the behavior
126 CHAPTER 6. EM ALGORITHMS FOR TEXT PROCESSING
of a Markov process is determined only by the state it is in at some time (not by how
it got to that state), αt (r) can always be computed in terms of the |S| values in αt−1 (·)
and the observation xt :
X
αt (r) = Br (xt ) · αt−1 (q) · Aq (r)
q∈S
We have now shown how to compute the probability of being in any state q at any time
t, having generated hx1 , x2 , . . . , xt i, with the forward algorithm. The probability of the
full sequence is the probability of being in time |x| and in any state, so the answer to
Question 1 can be computed simply by summing over α values at time |x| for all states:
X
Pr(x; θ) = α|x| (q)
q∈S
In summary, there are two ways of computing the probability that a sequence of obser-
vations x was generated by M: exhaustive enumeration with summing and the forward
algorithm. Figure 6.4 illustrates the two possibilities. The upper panel shows the naı̈ve
exhaustive approach, enumerating all 43 possible labels y0 of x and computing their joint
probability Pr(x, y0 ). Summing over all y0 , the marginal probability of x is found to be
0.00018. The lower panel shows the forward trellis, consisting of 4 × 3 cells. Summing
over the final column also yields 0.00018, the same result.
3. Given a set
using 3.theofforward
Given states
a 3.
set Given
ofanstates
algorithm
S, observation
a set S,ofanstates
(lower vocabulary
observation
panel). S, an observation and a series
vocabulary
O, of %a i.i.d.
vocabulary
O, and series
O, and of %a i.i.d.
series of % i.i.d.
y∗ = arg max Pr(x, y; θ)
observation sequences !x 1 , x
observation observation
sequences
2 , . . . , x!x", what
" sequences are
1 , x2 , . . . , x the
1, x
!x" ", y∈Yparameters
what
2 , . . .are
, x"the
", what
θ =
parameters
!A, B, that
are theπ"θparameters
= !A, B, π" θ=that!A, B, π" that
maximize themaximize
likelihoodthe of likelihood
the training
maximize data?
theoflikelihood
the training of thedata? training data?
3. Given a set of states S, an " !
observation vocabulary O, and a series of % i.i.d.
" " " ! " " !
observation sequences
θ∗ = arg max !x1θ, ∗x= 2 , .arg x" ",θi∗what
. . Pr(x
, max , y; are
= arg Pr(x
θ) max i , y; θ) Pr(xi , θy;=θ)!A, B, π" that
the parameters
maximize the likelihoodθ ofi=1 they∈Y trainingθ data?
i=1 y∈Y
θ
i=1 y∈Y
"
Given our definition
Given ourof definition
an HMM,
Given the
ourof ananswers
definition
HMM, ofto
thethe
an first"
answers
HMM, two!
the
to questions
the
answers toare
first two the infirst
principle
questions twoare
questions
in principle
are in principle
θ∗ = arg max Pr(xi , y; θ)
quite trivial to compute:
quite by iterating
trivial quite trivial over
to compute: to all stateby
bycompute:
iterating sequences
overiterating
all i=1
θ state
Y, the
over
y∈Y
probability
sequences
all stateY, that
sequences
the each
probability
Y, the that
probability
each that each
7 The 7 The
organization of this 7 The
section isofbased in part is
on ideassection
from
organization organization
this section ofbased
this in partLawrence
ison
based Rabiner’s
ideasinfrom HMMfrom
partLawrence
on ideas tutorial
Rabiner’s[64].
Lawrence
HMMRabiner’s
tutorial [64].
HMM tutorial [64].
7 The organization of this section is based in part on ideas from Lawrence Rabiner’s HMM tutorial [64].
upper panel upper
shows panel
the naive
upper
showsapproach,
panel
the naive
shows
enumerating
approach,
the naiveenumerating
all
approach,
43 possibleenumerating
all labels
43 possible
y! all
of x
labels
43and
possible
y! of xlabels
and y! of x and
computing their
computing
joint probability
their
computing
joint Pr(x,
probability
theiryjoint
!
). Summing
Pr(x,
probability
y ).over
!
Summing
Pr(x,
all yy, ).
! !
the
over
Summing
marginal
all y , the
!
over
proba-
marginal
all y , the
!
proba-
marginal proba-
bility of x isbility
foundoftox be
isbility
found
0.00018.
oftoxThe
be
is found
0.00018.
lowertopanel
The
be 0.00018.
shows
lower panel
the
Theforward
shows
lower panel
the
trellis,
forward
shows
consisting
the
trellis,
forward
consisting
trellis, consisting
of 4 × 3 cells.
of Summing
4 × 3 cells.
over
of Summing
4×the
3 cells.
finalover
column,
Summing
the final
also
over
column,
computes
the final
also
0.00018.
column,
computes also
0.00018.
computes 0.00018.
γ1 γ2 γ3 γ1 γ2 γ3 γ1 γ2 γ3
Figure 6.5: Computing the most likely state sequence that generated hJohn, might, washi un-
der the HMM given in Figure γ6.3 1 (q)using= the
πγq 1·(q)q (x=
Viterbi
B 1) π γq 1·(q)
Bq (x=
algorithm. 1 ) The
πq · B q (x1 )likely state sequence
most
bp (q) = −1
bp (q) = −1
bp (q) = −1
is highlighted in bold and could be recovered programmatically by following backpointers from
1 1 1
Thetherecursion
maximal
The isprobability
similar
recursion
The cell
to isthat in
of the
similar
recursion
the last
to forward
isthatcolumn
similar
of the to the
algorithm,
to forward
that first
ofexcept
the column
algorithm,
forward
rather (thicker
except
algorithm,
than rather arrows).
summingexcept
than rather
summingthan summing
over previousover
states,
previous
the over
maximum
states,
previous
thevalue
maximum
states,
of allthepossible
value
maximum
of trajectories
all possible
value ofinto
trajectories
all possible
state r at into
trajectories
time
state r at
into
time
state r at time
t is computed.t isNote
computed.
thattthe
isNote
computed.
back-pointer
that the Noteback-pointer
just
that
records
the back-pointer
just
the index
records of just
the index
originating
recordsof the state
index
originating
of thestate
originating state
– a separate–computation
a separate–computation
ais separate
not necessary.
computation
is not necessary. is not necessary.
γ1 (q) = πq · Bq (x1 )
bp1 (q) = −1
The recursion is similar to that of the forward algorithm, except rather than summing
over previous states, the maximum value of all possible trajectories into state r at time
t is computed. Note that the backpointer simply records the index of the originating
state—a separate computation is not necessary.
To compute the best sequence of states, y∗ , the state with the highest probability path
at time |x| is selected, and then the backpointers are followed, recursively, to construct
the rest of the sequence:
∗
y|x| = arg max γ|x| (q)
q∈S
∗
yt−1 = bpt (yt )
Figure 6.5 illustrates a Viterbi trellis, including backpointers that have been used to
compute the most likely state sequence.
6.2. HIDDEN MARKOV MODELS 129
ADJ
NN
Figure 6.6: A “fully observable” HMM training instance. The output sequence is at the top
of the figure, and the corresponding states and transitions are shown in the trellis below.
9 Since an HMM models sequences, its training data consists of a collection of example sequences.
130 CHAPTER 6. EM ALGORITHMS FOR TEXT PROCESSING
I(q) T (q → r) O(q ↑ o)
πq = P Aq (r) = P Bq (o) = P (6.2)
`= r I(r) N (q) = r0 T (q → r0 ) N (q) = o0 O(q ↑ o0 )
For example, to compute the emission parameters from state nn, we simply need to keep
track of the number of times the process is in state nn and what symbol it generates
at each of these times. Transition probabilities are computed similarly: to compute, for
example, the distribution Adet (·), that is, the probabilities of transitioning away from
state det, we count the number of times the process is in state det, and keep track
of what state the process transitioned into at the next time step. This counting and
normalizing be accomplished using the exact same counting and relative frequency al-
gorithms that we described in Section 3.3. Thus, in the fully observable case, parameter
estimation is not a new algorithm at all, but one we have seen before.
How should the model parameters be estimated when the state sequence is not
provided? It turns out that the update equations have the satisfying form where the
optimal parameter values for iteration i + 1 are expressed in terms of the expectations of
the counts referenced in the fully observed case, according to the posterior distribution
over the latent variables given the observations x and the parameters θ(i) :
As a result, the optimization problems (i.e., Equations 6.2) require completely indepen-
dent sets of statistics, which we will utilize later to facilitate efficient parallelization in
MapReduce.
How can the expectations in Equation 6.3 be understood? In the fully observed
training case, between every time step, there is exactly one transition taken and the
source and destination states are observable. By progressing through the Markov chain,
we can let each transition count as ‘1’, and we can accumulate the total number of times
each kind of transition was taken (by each kind, we simply mean the number of times
that one state follows another, for example, the number of times nn follows det). These
statistics can then in turn be used to compute the MLE for an ‘observable’ HMM, as
6.2. HIDDEN MARKOV MODELS 131
described above. However, when the transition sequence is not observable (as is most
often the case), we can instead imagine that at each time step, every possible transition
(there are |S|2 of them, and typically |S| is quite small) is taken, with a particular
probability. The probability used is the posterior probability of the transition, given the
model and an observation sequence (we describe how to compute this value below).
By summing over all the time steps in the training data, and using this probability as
the ‘count’ (rather than ‘1’ as in the observable case), we compute the expected count
of the number of times a particular transition was taken, given the training sequence.
Furthermore, since the training instances are statistically independent, the value of the
expectations can be computed by processing each training instance independently and
summing the results.
Similarly for the necessary emission counts (the number of times each symbol in
O was generated by each state in S), we assume that any state could have generated
the observation. We must therefore compute the probability of being in every state at
each time point, which is then the size of the emission ‘count’. By summing over all
time steps we compute the expected count of the number of times that a particular
state generated a particular symbol. These two sets of expectations, which are written
formally here, are sufficient to execute the M-step.
|x|
X
E[O(q ↑ o)] = Pr(yi = q|x; θ) · δ(xi , o) (6.4)
i=1
|x|−1
X
E[T (q → r)] = Pr(yi = q, yi+1 = r|x; θ) (6.5)
i=1
a b b c b
S1
S2
S3
α2(2) β3(2)
Figure 6.7: Using forward and backward probabilities to compute the posterior probability of
the dashed transition, given the observation sequence a b b c b. The shaded area on the left
corresponds to the forward probability α2 (s2 ), and the shaded area on the right corresponds to
the backward probability β3 (s2 ).
Equation 6.6 is the probability of being in state q at time i, given x, and the correctness
of the expression should be clear from the definitions of forward and backward proba-
bilities. The intuition for Equation 6.7, the probability of taking a particular transition
at a particular time, is also not complicated: it is the product of four conditionally
independent probabilities: the probability of getting to state q at time i (having gener-
ated the first part of the sequence), the probability of taking transition q → r (which
is specified in the parameters, θ), the probability of generating observation xi+1 from
state r (also specified in θ), and the probability of generating the rest of the sequence,
along any path. A visualization of the quantities used in computing this probability is
shown in Figure 6.7. In this illustration, we assume an HMM with S = {s1 , s2 , s3 } and
O = {a, b, c}.
The backward algorithm. Like the forward and Viterbi algorithms introduced
above to answer Questions 1 and 2, the backward algorithm uses dynamic program-
ming to incrementally compute βt (·). Its base case starts at time |x|, and is defined as
follows:
6.2. HIDDEN MARKOV MODELS 133
β|x| (q) = 1
To understand the intuition for this base case, keep in mind that since the backward
probabilities βt (·) are the probability of generating the remainder of the sequence after
time t (as well as being in some state), and since there is nothing left to generate after
time |x|, the probability must be 1. The recursion is defined as follows:
X
βt (q) = βt+1 (r) · Aq (r) · Br (xt+1 )
r∈S
Unlike the forward and Viterbi algorithms, the backward algorithm is computed from
right to left and makes no reference to the start probabilities, π.
This can be useful, since it provides a way of constraining the structures of the Markov
model; however, one must be aware of this behavior.
Another pitfall to avoid when implementing HMMs is arithmetic underflow.
HMMs typically define a massive number of sequences, and so the probability of any one
of them is often vanishingly small—so small that they often underflow standard floating
point representations. A very common solution to this problem is to represent prob-
abilities using their logarithms. Note that expected counts do not typically have this
problem and can be represented using normal floating point numbers. See Section 5.4
for additional discussion on working with log probabilities.
6.3 EM IN MAPREDUCE
Expectation maximization algorithms fit quite naturally into the MapReduce program-
ming model. Although the model being optimized determines the details of the required
computations, MapReduce implementations of EM algorithms share a number of char-
acteristics:
• Each iteration of EM is one MapReduce job.
• A controlling process (i.e., driver program) spawns the MapReduce jobs, keeps
track of the number of iterations and convergence criteria.
• Model parameters θ(i) , which are static for the duration of the MapReduce job,
are loaded by each mapper from HDFS or other data provider (e.g., a distributed
key-value store).
• Mappers map over independent training instances, computing partial latent vari-
able posteriors (or summary statistics, such as expected counts).
• Reducers sum together the required training statistics and solve one or more of
the M-step optimization problems.
• Combiners, which sum together the training statistics, are often quite effective at
reducing the amount of data that must be written to disk.
The degree of parallelization that can be attained depends on the statistical indepen-
dence assumed in the model and in the derived quantities required to solve the opti-
mization problems in the M-step. Since parameters are estimated from a collection of
samples that are assumed to be i.i.d., the E-step can generally be parallelized effectively
since every training instance can be processed independently of the others. In the limit,
in fact, each independent training instance could be processed by a separate mapper!10
10 Although the wisdom of doing this is questionable, given that the startup costs associated with individual map
tasks in Hadoop may be considerable.
6.3. EM IN MAPREDUCE 135
Reducers, however, must aggregate the statistics necessary to solve the optimiza-
tion problems as required by the model. The degree to which these may be solved
independently depends on the structure of the model, and this constrains the number
of reducers that may be used. Fortunately, many common models (such as HMMs) re-
quire solving several independent optimization problems in the M-step. In this situation,
a number of reducers may be run in parallel. Still, it is possible that in the worst case,
the M-step optimization problem will not decompose into independent subproblems,
making it necessary to use a single reducer.
HMM training mapper. The pseudo-code for the HMM training mapper is given
in Figure 6.8. The input consists of key-value pairs with a unique id as the key and
a training instance (e.g., a sentence) as the value. For each training instance, 2n + 1
stripes are emitted with unique keys, and every training instance emits the same set
of keys. Each unique key corresponds to one of the independent optimization problems
that will be solved in the M-step. The outputs are:
136 CHAPTER 6. EM ALGORITHMS FOR TEXT PROCESSING
1: class Mapper
2: method Initialize(integer iteration)
3: hS, Oi ← ReadModel
4: θ ← hA, B, πi ← ReadModelParams(iteration)
5: method Map(sample id, sequence x)
6: α ← Forward(x, θ) . cf. Section 6.2.2
7: β ← Backward(x, θ) . cf. Section 6.2.4
8: I ← new AssociativeArray . Initial state expectations
9: for all q ∈ S do . Loop over states
10: I{q} ← α1 (q) · β1 (q)
11: O ← new AssociativeArray of AssociativeArray . Emissions
12: for t = 1 to |x| do . Loop over observations
13: for all q ∈ S do . Loop over states
14: O{q}{xt } ← O{q}{xt } + αt (q) · βt (q)
15: t←t+1
16: T ← new AssociativeArray of AssociativeArray . Transitions
17: for t = 1 to |x| − 1 do . Loop over observations
18: for all q ∈ S do . Loop over states
19: for all r ∈ S do . Loop over states
20: T {q}{r} ← T {q}{r} + αt (q) · Aq (r) · Br (xt+1 ) · βt+1 (r)
21: t←t+1
22: Emit(string ‘initial ’, stripe I)
23: for all q ∈ S do . Loop over states
24: Emit(string ‘emit from ’ + q, stripe O{q})
25: Emit(string ‘transit from ’ + q, stripe T {q})
Figure 6.8: Mapper pseudo-code for training hidden Markov models using EM. The mappers
map over training instances (i.e., sequences of observations xi ) and generate the expected counts
of initial states, emissions, and transitions taken to generate the sequence.
6.3. EM IN MAPREDUCE 137
1: class Combiner
2: method Combine(string t, stripes [C1 , C2 , . . .])
3: Cf ← new AssociativeArray
4: for all stripe C ∈ stripes [C1 , C2 , . . .] do
5: Sum(Cf , C)
6: Emit(string t, stripe Cf )
1: class Reducer
2: method Reduce(string t, stripes [C1 , C2 , . . .])
3: Cf ← new AssociativeArray
4: for all stripe C ∈ stripes [C1 , C2 , . . .] do
5: Sum(Cf , C)
6: z←0
7: for all hk, vi ∈ Cf do
8: z ←z+v
9: Pf ← new AssociativeArray . Final parameters vector
10: for all hk, vi ∈ Cf do
11: Pf {k} ← v/z
12: Emit(string t, stripe Pf )
Figure 6.9: Combiner and reducer pseudo-code for training hidden Markov models using EM.
The HMMs considered in this book are fully parameterized by multinomial distributions, so
reducers do not require special logic to handle different types of model parameters (since they
are all of the same type).
138 CHAPTER 6. EM ALGORITHMS FOR TEXT PROCESSING
1. the probabilities that the unobserved Markov process begins in each state q, with
a unique key designating that the values are initial state counts;
2. the expected number of times that state q generated each emission symbol o (the
set of emission symbols included will be just those found in each training instance
x), with a key indicating that the associated value is a set of emission counts from
state q; and
3. the expected number of times state q transitions to each state r, with a key indi-
cating that the associated value is a set of transition counts from state q.
HMM training reducer. The reducer for one iteration of HMM training, shown
together with an optional combiner in Figure 6.9, aggregates the count collections as-
sociated with each key by summing them. When the values for each key have been
completely aggregated, the associative array contains all of the statistics necessary to
compute a subset of the parameters for the next EM iteration. The optimal parameter
settings for the following iteration are computed simply by computing the relative fre-
quency of each event with respect to its expected count at the current iteration. The
new computed parameters are emitted from the reducer and written to HDFS. Note
that they will be spread across 2 · |S| + 1 keys, representing initial state probabilities
π, transition probabilities Aq for each state q, and emission probabilities Bq for each
state q.
With the statistical approach, translation systems can be developed cheaply and
quickly for any language pair, as long as there is sufficient training data available.
Furthermore, improvements in learning algorithms and statistical modeling can yield
benefits in many translation pairs at once, rather than being specific to individual
language pairs. Thus, SMT, like many other topics we are considering in this book,
is an attempt to leverage the vast quantities of textual data that is available to solve
problems that would otherwise require considerable manual effort to encode specialized
knowledge. Since the advent of statistical approaches to translation, the field has grown
tremendously and numerous statistical models of translation have been developed, with
many incorporating quite specialized knowledge about the behavior of natural language
as biases in their learning algorithms.
n
Y
Pr(w1n ) = Pr(w1 ) Pr(w2 |w1 ) Pr(w3 |w12 ) . . . Pr(wn |w1n−1 ) = Pr(wk |w1k−1 ) (6.8)
k=1
Due to the extremely large number of parameters involved in estimating such a model
directly, it is customary to make the Markov assumption, that the sequence histories
only depend on prior local context. That is, an n-gram language model is equivalent to
a (n − 1)th-order Markov model. Thus, we can approximate P (wk |w1k−1 ) as follows:
The probabilities used in computing Pr(w1n ) based on an n-gram language model are
generally estimated from a monolingual corpus of target language text. Since only target
6.4. CASE STUDY: WORD ALIGNMENT FOR STATISTICAL MACHINE TRANSLATION 141
(vi, i saw)
i saw the small table
(la mesa pequeña, the small table)
vi la mesa pequeña …
Parallel Sentences
Decoder
maria no daba una bofetada a la bruja verde mary did not slap the green witch
Foreign Input Sentence English Output Sentence
Figure 6.10: The standard phrase-based machine translation architecture. The translation
model is constructed with phrases extracted from a word-aligned parallel corpus. The language
model is estimated from a monolingual corpus. Both serve as input to the decoder, which
performs the actual translation.
no slap to the
the
Figure 6.11: Translation coverage of the sentence Maria no dio una bofetada a la bruja verde
by a phrase-based model. The best possible translation path is indicated with a dashed line.
142 CHAPTER 6. EM ALGORITHMS FOR TEXT PROCESSING
language text is necessary (without any additional annotation), language modeling has
been well served by large-data approaches that take advantage of the vast quantities of
text available on the web.
To translate an input sentence f, the phrase-based decoder creates a matrix of all
translation possibilities of all substrings in the input string, as an example illustrates in
Figure 6.11. A sequence of phrase pairs is selected such that each word in f is translated
exactly once.13 The decoder seeks to find the translation that maximizes the product of
the translation probabilities of the phrases used and the language model probability of
the resulting string in the target language. Because the phrase translation probabilities
are independent of each other and the Markov assumption made in the language model,
this may be done efficiently using dynamic programming. For a detailed introduction
to phrase-based decoding, we refer the reader to a recent textbook by Koehn [85].
13 Thephrases may not necessarily be selected in a strict left-to-right order. Being able to vary the order of the
phrases used is necessary since languages may express the same ideas using different word orders.
6.4. CASE STUDY: WORD ALIGNMENT FOR STATISTICAL MACHINE TRANSLATION 143
and eventually disappeared with sufficient data. Furthermore, with stupid backoff it
was possible to train a language model on more data than was feasible with Kneser-
Ney smoothing. Applying this language model to a machine translation task yielded
better results than a (smaller) language model trained with Kneser-Ney smoothing.
The role of the language model in statistical machine translation is to select
fluent, grammatical translations from a large hypothesis space: the more training data a
language model has access to, the better its description of relevant language phenomena
and hence its ability to select good translations. Once again, large data triumphs! For
more information about estimating language models using MapReduce, we refer the
reader to a forthcoming book from Morgan & Claypool [26].
|e|
Y
Pr(e, a|f) = Pr(a|f, e) × Pr(ei |fai )
| {z } i=1
Alignment probability | {z }
Lexical probability
Since we have parallel corpora consisting of only hf, ei pairs, we can learn the parame-
ters for this model using EM and treating a as a latent variable. However, to combat
14 Inthe original presentation of statistical lexical translation models, a special null word is added to the source
sentences, which permits words to be inserted ‘out of nowhere’. Since this does not change any of the important
details of training, we omit it from our presentation for simplicity.
144 CHAPTER 6. EM ALGORITHMS FOR TEXT PROCESSING
data sparsity in the alignment probability, we must make some further simplifying as-
sumptions. By letting the probability of an alignment depend only on the position of
the previous aligned word we capture a valuable insight (namely, words that are nearby
in the source language will tend to be nearby in the target language), and our model
acquires the structure of an HMM [150]:
|e| |e|
Y Y
Pr(e, a|f) = Pr(ai |ai−1 ) × Pr(ei |fai )
i=1 i=1
| {z } | {z }
Transition probability Emission probability
This model can be trained using the forward-backward algorithm described in the pre-
vious section, summing over all settings of a, and the best alignment for a sentence pair
can be found using the Viterbi algorithm.
To properly initialize this HMM, it is conventional to further simplify the align-
ment probability model, and use this simpler model to learn initial lexical translation
(emission) parameters for the HMM. The favored simplification is to assert that all
alignments are uniformly probable:
|e|
1 Y
Pr(e, a|f) = |e|
× Pr(ei |fai )
|f| i=1
6.4.4 EXPERIMENTS
How well does a MapReduce word aligner for statistical machine translation perform?
We describe previously-published results [54] that compared a Java-based Hadoop im-
plementation against a highly optimized word aligner called Giza++ [112], which was
written in C++ and designed to run efficiently on a single core. We compared the train-
ing time of Giza++ and our aligner on a Hadoop cluster with 19 slave nodes, each with
two single-core processors and two disks (38 cores total).
Figure 6.12 shows the performance of Giza++ in terms of the running time of a
single EM iteration for both Model 1 and the HMM alignment model as a function of
the number of training pairs. Both axes in the figure are on a log scale, but the ticks
6.4. CASE STUDY: WORD ALIGNMENT FOR STATISTICAL MACHINE TRANSLATION 145
on the y-axis are aligned with ‘meaningful’ time intervals rather than exact orders of
magnitude. There are three things to note. First, the running time scales linearly with
the size of the training data. Second, the HMM is a constant factor slower than Model 1.
Third, the alignment process is quite slow as the size of the training data grows—at one
million sentences, a single iteration takes over three hours to complete! Five iterations
are generally necessary to train the models, which means that full training takes the
better part of a day.
In Figure 6.13 we plot the running time of our MapReduce implementation run-
ning on the 38-core cluster described above. For reference, we plot points indicating
what 1/38 of the running time of the Giza++ iterations would be at each data size,
which gives a rough indication of what an ‘ideal’ parallelization could achieve, assum-
ing that there was no overhead associated with distributing computation across these
machines. Three things may be observed in the results. First, as the amount of data
increases, the relative cost of the overhead associated with distributing data, marshal-
ing and aggregating counts, decreases. At one million sentence pairs of training data,
the HMM alignment iterations begin to approach optimal runtime efficiency. Second,
Model 1, which we observe is light on computation, does not approach the theoretical
performance of an ideal parallelization, and in fact, has almost the same running time
as the HMM alignment algorithm. We conclude that the overhead associated with dis-
tributing and aggregating data is significant compared to the Model 1 computations,
although a comparison with Figure 6.12 indicates that the MapReduce implementation
is still substantially faster than the single core implementation, at least once a certain
training data size is reached. Finally, we note that, in comparison to the running times
of the single-core implementation, at large data sizes, there is a significant advantage
to using the distributed implementation, even of Model 1.
Although these results do confound several variables (Java vs. C++ performance,
memory usage patterns), it is reasonable to expect that the confounds would tend to
make the single-core system’s performance appear relatively better than the MapReduce
system (which is, of course, the opposite pattern from what we actually observe). Fur-
thermore, these results show that when computation is distributed over a cluster of
many machines, even an unsophisticated implementation of the HMM aligner could
compete favorably with a highly optimized single-core system whose performance is
well-known to many people in the MT research community.
Why are these results important? Perhaps the most significant reason is that
the quantity of parallel data that is available to train statistical machine translation
models is ever increasing, and as is the case with so many problems we have encountered,
more data leads to improvements in translation quality [54]. Recently a corpus of one
billion words of French-English data was mined automatically from the web and released
146 CHAPTER 6. EM ALGORITHMS FOR TEXT PROCESSING
Model 1
HMM
Average iteration latency (seconds)
3 hrs
60 min
20 min
5 min
90 s
30 s
10 s
3s
Figure 6.12: Running times of Giza++ (baseline single-core system) for Model 1 and HMM
training iterations at various corpus sizes.
5 min
90 s
30 s
10 s
3s
Figure 6.13: Running times of our MapReduce implementation of Model 1 and HMM training
iterations at various corpus sizes. For reference, 1/38 running times of the Giza++ models are
shown.
6.5. EM-LIKE ALGORITHMS 147
publicly [33]. Single-core solutions to model construction simply cannot keep pace with
15
the amount of translated data that is constantly being produced. Fortunately, several
independent researchers have shown that existing modeling algorithms can be expressed
naturally and effectively using MapReduce, which means that we can take advantage
of this data. Furthermore, the results presented here show that even at data sizes
that may be tractable on single machines, significant performance improvements are
attainable using MapReduce implementations. This improvement reduces experimental
turnaround times, which allows researchers to more quickly explore the solution space—
which will, we hope, lead to rapid new developments in statistical machine translation.
For the reader interested in statistical machine translation, there is an open source
Hadoop-based MapReduce implementation of a training pipeline for phrase-based trans-
lation that includes word alignment, phrase extraction, and phrase scoring [56].
∇F (θ∗ ) = 0
The parameter η (i) > 0 is a learning rate which indicates how quickly the algorithm
moves along the gradient during iteration i. Provided this value is small enough that
F decreases, this strategy will find a local minimum of F . However, while simple, this
update strategy may converge slowly, and proper selection of η is non-trivial. More
sophisticated algorithms perform updates that are informed by approximations of the
second derivative, which are estimated by successive evaluations of ∇F (θ), and can
converge much more rapidly [96].
• The objective should decompose linearly across training instances. This implies
that the gradient also decomposes linearly, and therefore mappers can process
input data in parallel. The values they emit are pairs hF (θ), ∇F (θ)i, which are
linear components of the objective and gradient.
6.5. EM-LIKE ALGORITHMS 149
• Whether more than one reducer can run in parallel depends on the specific opti-
mization algorithm being used. Some, like the trivial algorithm of Equation 6.12
treat the dimensions of θ independently, whereas many are sensitive to global
properties of ∇F (θ). In the latter case, parallelization across multiple reducers is
non-trivial.
• Reducer(s) sum the component objective/gradient pairs, compute the total objec-
tive and gradient, run the optimization algorithm, and emit θ(i+1) .
• Many optimization algorithms are stateful and must persist their state between
optimization iterations. This may either be emitted together with θ(i+1) or written
to the distributed file system as a side effect of the reducer. Such external side
effects must be handled carefully; refer to Section 2.2 for a discussion.
In this expression, Hi are real-valued functions sensitive to features of the input and
labeling. The parameters of the model is selected so as to minimize the negative condi-
tional log likelihood of a set of training instances hhx, yi1 , hx, yi2 , . . .i, which we assume
to be i.i.d.:
X
F (θ) = − log Pr(y|x; θ) (6.13)
hx,yi
θ∗ = arg min F (θ) (6.14)
θ
150 CHAPTER 6. EM ALGORITHMS FOR TEXT PROCESSING
As Equation 6.13 makes clear, the objective decomposes linearly across training in-
stances, meaning it can be optimized quite well in MapReduce. The gradient derivative
of F with respect to θi can be shown to have the following form [141]:16
∂F X
(θ) = Hi (x, y) − EPr(y0 |x;θ) [Hi (x, y0 )]
∂θi hx,yi
The expectation in the second part of the gradient’s expression can be computed using a
variety of techniques. However, as we saw with EM, when very large event spaces are be-
ing modeled, as is the case with sequence labeling, enumerating all possible values y can
become computationally intractable. And, as was the case with HMMs, independence
assumptions can be used to enable efficient computation using dynamic programming.
In fact, the forward-backward algorithm introduced in Section 6.2.4 can, with only min-
imal modification, be used to compute the expectation EPr(y0 |x;θ) [Hi (x, y0 )] needed in
CRF sequence models, as long as the feature functions respect the same Markov as-
sumption that is made in HMMs. For more information about inference in CRFs using
the forward-backward algorithm, we refer the reader to Sha et al. [140].
As we saw in the previous section, MapReduce offers significant speedups when
training iterations require running the forward-backward algorithm. The same pattern
of results holds when training linear CRFs.
large datasets in parallel, provides researchers with an effective strategy for developing
increasingly-effective applications.
Since EM algorithms are relatively computationally expensive, even for small
amounts of data, this led us to consider how related supervised learning models (which
typically have much less training data available), can also be implemented in Map-
Reduce. The discussion demonstrates that not only does MapReduce provide a means
for coping with ever-increasing amounts of data, but it is also useful for parallelizing
expensive computations. Although MapReduce has been designed with mostly data-
intensive applications in mind, the ability to leverage clusters of commodity hardware
to parallelize computationally-expensive algorithms is an important use case.
Additional Readings. Because of its ability to leverage large amounts of training
data, machine learning is an attractive problem for MapReduce and an area of active
research. Chu et al. [37] presented general formulations of a variety of machine learning
problems, focusing on a normal form for expressing a variety of machine learning algo-
rithms in MapReduce. The Apache Mahout project is an open-source implementation
of these and other learning algorithms,17 and it is also the subject of a forthcoming
book [116]. Issues associated with a MapReduce implementation of latent Dirichlet
allocation (LDA), which is another important unsupervised learning technique, with
certain similarities to EM, have been explored by Wang et al. [151].
17 http://lucene.apache.org/mahout/
152
CHAPTER 7
Closing Remarks
The need to process enormous quantities of data has never been greater. Not only
are terabyte- and petabyte-scale datasets rapidly becoming commonplace, but there
is consensus that great value lies buried in them, waiting to be unlocked by the right
computational tools. In the commercial sphere, business intelligence—driven by the
ability to gather data from a dizzying array of sources—promises to help organizations
better understand their customers and the marketplace, hopefully leading to better
business decisions and competitive advantages. For engineers building information pro-
cessing tools and applications, larger datasets lead to more effective algorithms for a
wide range of tasks, from machine translation to spam detection. In the natural and
physical sciences, the ability to analyze massive amounts of data may provide the key
to unlocking the secrets of the cosmos or the mysteries of life.
In the preceding chapters, we have shown how MapReduce can be exploited to
solve a variety of problems related to text processing at scales that would have been
unthinkable a few years ago. However, no tool—no matter how powerful or flexible—
can be perfectly adapted to every task, so it is only fair to discuss the limitations of
the MapReduce programming model and survey alternatives. Section 7.1 covers online
learning algorithms and Monte Carlo simulations, which are examples of algorithms
that require maintaining global state. As we have seen, this is difficult to accomplish
in MapReduce. Section 7.2 discusses alternative programming models, and the book
concludes in Section 7.3.
updates are not made until the full evidence of the training data has been weighed
against the model. An earlier update would seem, in some sense, to be hasty. However,
it is generally the case that more frequent updates can lead to more rapid convergence
of the model (in terms of number of training instances processed), even if those updates
are made by considering less data [24]. Thinking in terms of gradient optimization (see
Section 6.5), online learning algorithms can be understood as computing an approx-
imation of the true gradient, using only a few training instances. Although only an
approximation, the gradient computed from a small subset of training instances is of-
ten quite reasonable, and the aggregate behavior of multiple updates tends to even out
errors that are made. In the limit, updates can be made after every training instance.
Unfortunately, implementing online learning algorithms in MapReduce is problem-
atic. The model parameters in a learning algorithm can be viewed as shared global state,
which must be updated as the model is evaluated against training data. All processes
performing the evaluation (presumably the mappers) must have access to this state.
In a batch learner, where updates occur in one or more reducers (or, alternatively, in
the driver code), synchronization of this resource is enforced by the MapReduce frame-
work. However, with online learning, these updates must occur after processing smaller
numbers of instances. This means that the framework must be altered to support faster
processing of smaller datasets, which goes against the design choices of most existing
MapReduce implementations. Since MapReduce was specifically optimized for batch
operations over large amounts of data, such a style of computation would likely result
in inefficient use of resources. In Hadoop, for example, map and reduce tasks have con-
siderable startup costs. This is acceptable because in most circumstances, this cost is
amortized over the processing of many key-value pairs. However, for small datasets,
these high startup costs become intolerable. An alternative is to abandon shared global
state and run independent instances of the training algorithm in parallel (on different
portions of the data). A final solution is then arrived at by merging individual results.
Experiments, however, show that the merged solution is inferior to the output of running
the training algorithm on the entire dataset [52].
A related difficulty occurs when running what are called Monte Carlo simula-
tions, which are used to perform inference in probabilistic models where evaluating or
representing the model exactly is impossible. The basic idea is quite simple: samples
are drawn from the random variables in the model to simulate its behavior, and then
simple frequency statistics are computed over the samples. This sort of inference is par-
ticularly useful when dealing with so-called nonparametric models, which are models
whose structure is not specified in advance, but is rather inferred from training data.
For an illustration, imagine learning a hidden Markov model, but inferring the num-
ber of states, rather than having them specified. Being able to parallelize Monte Carlo
simulations would be tremendously valuable, particularly for unsupervised learning ap-
154 CHAPTER 7. CLOSING REMARKS
plications where they have been found to be far more effective than EM-based learning
(which requires specifying the model). Although recent work [10] has shown that the
delays in synchronizing sample statistics due to parallel implementations do not neces-
sarily damage the inference, MapReduce offers no natural mechanism for managing the
global shared state that would be required for such an implementation.
The problem of global state is sufficiently pervasive that there has been substan-
tial work on solutions. One approach is to build a distributed datastore capable of
maintaining the global state. However, such a system would need to be highly scal-
able to be used in conjunction with MapReduce. Google’s BigTable [34], which is a
sparse, distributed, persistent multidimensional sorted map built on top of GFS, fits
the bill, and has been used in exactly this manner. Amazon’s Dynamo [48], which is a
distributed key-value store (with a very different architecture), might also be useful in
this respect, although it wasn’t originally designed with such an application in mind.
Unfortunately, it is unclear if the open-source implementations of these two systems
(HBase and Cassandra, respectively) are sufficiently mature to handle the low-latency
and high-throughput demands of maintaining global state in the context of massively
distributed processing (but recent benchmarks are encouraging [40]).
are used to transport partial results between vertices, and can be realized using files,
TCP pipes, or shared memory.
Another system worth mentioning is Pregel [98], which implements a program-
ming model inspired by Valiant’s Bulk Synchronous Parallel (BSP) model [148]. Pregel
was specifically designed for large-scale graph algorithms, but unfortunately there are
few published details at present. However, a longer description is anticipated in a forth-
coming paper [99].
What is the significance of these developments? The power of MapReduce derives
from providing an abstraction that allows developers to harness the power of large
clusters. As anyone who has taken an introductory computer science course would know,
abstractions manage complexity by hiding details and presenting well-defined behaviors
to users of those abstractions. This process makes certain tasks easier, but others more
difficult, if not impossible. MapReduce is certainly no exception to this generalization,
and one of the goals of this book has been to give the reader a better understanding of
what’s easy to do in MapReduce and what its limitations are. But of course, this begs
the obvious question: What other abstractions are available in the massively-distributed
datacenter environment? Are there more appropriate computational models that would
allow us to tackle classes of problems that are difficult for MapReduce?
Dryad and Pregel are alternative answers to these questions. They share in pro-
viding an abstraction for large-scale distributed computations, separating the what from
the how of computation and isolating the developer from the details of concurrent pro-
gramming. They differ, however, in how distributed computations are conceptualized:
functional-style programming, arbitrary dataflows, or BSP. These conceptions represent
different tradeoffs between simplicity and expressivity: for example, Dryad is more flex-
ible than MapReduce, and in fact, MapReduce can be trivially implemented in Dryad.
However, it remains unclear, at least at present, which approach is more appropriate
for different classes of applications. Looking forward, we can certainly expect the de-
velopment of new models and a better understanding of existing ones. MapReduce is
not the end, and perhaps not even the best. It is merely the first of many approaches
to harness large-scaled distributed computing resources.
Even within the Hadoop/MapReduce ecosystem, we have already observed the
development of alternative approaches for expressing distributed computations. For
example, there is a proposal to add a third merge phase after map and reduce to
better support relational operations [36]. Pig [114], which was inspired by Google’s
Sawzall [122], can be described as a data analytics platform that provides a lightweight
scripting language for manipulating large datasets. Although Pig scripts (in a language
called Pig Latin) are ultimately converted into Hadoop jobs by Pig’s execution engine,
constructs in the language allow developers to specify data transformations (filtering,
joining, grouping, etc.) at a much higher level. Similarly, Hive [68], another open-source
156 CHAPTER 7. CLOSING REMARKS
project, provides an abstraction on top of Hadoop that allows users to issue SQL queries
against large relational datasets stored in HDFS. Hive queries (in HiveQL) “compile
down” to Hadoop jobs by the Hive query engine. Therefore, the system provides a data
analysis tool for users who are already comfortable with relational databases, while
simultaneously taking advantage of Hadoop’s data processing capabilities.
Bibliography
[1] Azza Abouzeid, Kamil Bajda-Pawlikowski, Daniel Abadi, Avi Silberschatz, and
Alexander Rasin. HadoopDB: An architectural hybrid of MapReduce and DBMS
technologies for analytical workloads. In Proceedings of the 35th International
Conference on Very Large Data Base (VLDB 2009), pages 922–933, Lyon, France,
2009.
[2] Réka Albert and Albert-László Barabási. Statistical mechanics of complex net-
works. Reviews of Modern Physics, 74:47–97, 2002.
[3] Noga Alon, Yossi Matias, and Mario Szegedy. The space complexity of approxi-
mating the frequency moments. In Proceedings of the 28th Annual ACM Sympo-
sium on Theory of Computing (STOC ’96), pages 20–29, Philadelphia, Pennsyl-
vania, 1996.
[4] Peter Alvaro, Tyson Condie, Neil Conway, Khaled Elmeleegy, Joseph M. Heller-
stein, and Russell C. Sears. BOOM: Data-centric programming in the datacenter.
Technical Report UCB/EECS-2009-98, Electrical Engineering and Computer Sci-
ences, University of California at Berkeley, 2009.
[5] Gene Amdahl. Validity of the single processor approach to achieving large-scale
computing capabilities. In Proceedings of the AFIPS Spring Joint Computer Con-
ference, pages 483–485, 1967.
[6] Rajagopal Ananthanarayanan, Karan Gupta, Prashant Pandey, Himabindu
Pucha, Prasenjit Sarkar, Mansi Shah, and Renu Tewari. Cloud analytics: Do
we really need to reinvent the storage stack? In Proceedings of the 2009 Workshop
on Hot Topics in Cloud Computing (HotCloud 09), San Diego, California, 2009.
[7] Thomas Anderson, Michael Dahlin, Jeanna Neefe, David Patterson, Drew Roselli,
and Randolph Wang. Serverless network file systems. In Proceedings of the 15th
ACM Symposium on Operating Systems Principles (SOSP 1995), pages 109–126,
Copper Mountain Resort, Colorado, 1995.
[8] Vo Ngoc Anh and Alistair Moffat. Inverted index compression using word-aligned
binary codes. Information Retrieval, 8(1):151–166, 2005.
[9] Michael Armbrust, Armando Fox, Rean Griffith, Anthony D. Joseph, Randy H.
Katz, Andrew Konwinski, Gunho Lee, David A. Patterson, Ariel Rabkin, Ion
158 CHAPTER 7. CLOSING REMARKS
Stoica, and Matei Zaharia. Above the clouds: A Berkeley view of cloud comput-
ing. Technical Report UCB/EECS-2009-28, Electrical Engineering and Computer
Sciences, University of California at Berkeley, 2009.
[10] Arthur Asuncion, Padhraic Smyth, and Max Welling. Asynchronous distributed
learning of topic models. In Advances in Neural Information Processing Systems
21 (NIPS 2008), pages 81–88, Vancouver, British Columbia, Canada, 2008.
[11] Ricardo Baeza-Yates, Carlos Castillo, Flavio Junqueira, Vassilis Plachouras, and
Fabrizio Silvestri. Challenges on distributed web retrieval. In Proceedings of the
IEEE 23rd International Conference on Data Engineering (ICDE 2007), pages
6–20, Istanbul, Turkey, 2007.
[12] Ricardo Baeza-Yates, Carlos Castillo, and Vicente López. PageRank increase un-
der different collusion topologies. In Proceedings of the First International Work-
shop on Adversarial Information Retrieval on the Web (AIRWeb 2005), pages
17–24, Chiba, Japan, 2005.
[13] Ricardo Baeza-Yates, Aristides Gionis, Flavio Junqueira, Vanessa Murdock, Vas-
silis Plachouras, and Fabrizio Silvestri. The impact of caching on search engines.
In Proceedings of the 30th Annual International ACM SIGIR Conference on Re-
search and Development in Information Retrieval (SIGIR 2007), pages 183–190,
Amsterdam, The Netherlands, 2007.
[14] Michele Banko and Eric Brill. Scaling to very very large corpora for natural
language disambiguation. In Proceedings of the 39th Annual Meeting of the Asso-
ciation for Computational Linguistics (ACL 2001), pages 26–33, Toulouse, France,
2001.
[15] Paul Barham, Boris Dragovic, Keir Fraser, Steven Hand, Tim Harris, Alex Ho,
Rolf Neugebauer, Ian Pratt, and Andrew Warfield. Xen and the art of virtualiza-
tion. In Proceedings of the 19th ACM Symposium on Operating Systems Principles
(SOSP 2003), pages 164–177, Bolton Landing, New York, 2003.
[16] Luiz André Barroso, Jeffrey Dean, and Urs Hölzle. Web search for a planet: The
Google cluster architecture. IEEE Micro, 23(2):22–28, 2003.
[17] Luiz André Barroso and Urs Hölzle. The case for energy-proportional computing.
Computer, 40(12):33–37, 2007.
[18] Luiz André Barroso and Urs Hölzle. The Datacenter as a Computer: An Introduc-
tion to the Design of Warehouse-Scale Machines. Morgan & Claypool Publishers,
2009.
7.3. MAPREDUCE AND BEYOND 159
[19] Jacek Becla, Andrew Hanushevsky, Sergei Nikolaev, Ghaleb Abdulla, Alex Szalay,
Maria Nieto-Santisteban, Ani Thakar, and Jim Gray. Designing a multi-petabyte
database for LSST. SLAC Publications SLAC-PUB-12292, Stanford Linear Ac-
celerator Center, May 2006.
[20] Jacek Becla and Daniel L. Wang. Lessons learned from managing a petabyte.
In Proceedings of the Second Biennial Conference on Innovative Data Systems
Research (CIDR 2005), Asilomar, California, 2005.
[21] Gordon Bell, Tony Hey, and Alex Szalay. Beyond the data deluge. Science,
323(5919):1297–1298, 2009.
[22] Monica Bianchini, Marco Gori, and Franco Scarselli. Inside PageRank. ACM
Transactions on Internet Technology, 5(1):92–128, 2005.
[23] Jorge Luis Borges. Collected Fictions (translated by Andrew Hurley). Penguin,
1999.
[24] Léon Bottou. Stochastic learning. In Olivier Bousquet and Ulrike von Luxburg,
editors, Advanced Lectures on Machine Learning, Lecture Notes in Artificial In-
telligence, LNAI 3176, pages 146–168. Springer Verlag, Berlin, 2004.
[25] Thorsten Brants, Ashok C. Popat, Peng Xu, Franz J. Och, and Jeffrey Dean.
Large language models in machine translation. In Proceedings of the 2007 Joint
Conference on Empirical Methods in Natural Language Processing and Computa-
tional Natural Language Learning, pages 858–867, Prague, Czech Republic, 2007.
[26] Thorsten Brants and Peng Xu. Distributed Language Models. Morgan & Claypool
Publishers, 2010.
[27] Eric Brill, Jimmy Lin, Michele Banko, Susan Dumais, and Andrew Ng. Data-
intensive question answering. In Proceedings of the Tenth Text REtrieval Confer-
ence (TREC 2001), pages 393–400, Gaithersburg, Maryland, 2001.
[29] Peter F. Brown, Vincent J. Della Pietra, Stephen A. Della Pietra, and Robert L.
Mercer. The mathematics of statistical machine translation: Parameter estima-
tion. Computational Linguistics, 19(2):263–311, 1993.
[31] Rajkumar Buyya, Chee Shin Yeo, Srikumar Venugopal, James Broberg, and Ivona
Brandic. Cloud computing and emerging IT platforms: Vision, hype, and reality
for delivering computing as the 5th utility. Future Generation Computer Systems,
25(6):599–616, 2009.
[32] Luis-Felipe Cabrera and Darrell D. E. Long. Swift: Using distributed disk striping
to provide high I/O data rates. Computer Systems, 4(4):405–436, 1991.
[33] Chris Callison-Burch, Philipp Koehn, Christof Monz, and Josh Schroeder. Find-
ings of the 2009 workshop on statistical machine translation. In Proceedings of the
Fourth Workshop on Statistical Machine Translation (StatMT ’09), pages 1–28,
Athens, Greece, 2009.
[34] Fay Chang, Jeffrey Dean, Sanjay Ghemawat, Wilson C. Hsieh, Deborah A. Wal-
lach, Michael Burrows, Tushar Chandra, Andrew Fikes, and Robert Gruber.
Bigtable: A distributed storage system for structured data. In Proceedings of the
7th Symposium on Operating System Design and Implementation (OSDI 2006),
pages 205–218, Seattle, Washington, 2006.
[35] Stanley F. Chen and Joshua Goodman. An empirical study of smoothing tech-
niques for language modeling. In Proceedings of the 34th Annual Meeting of
the Association for Computational Linguistics (ACL 1996), pages 310–318, Santa
Cruz, California, 1996.
[36] Hung chih Yang, Ali Dasdan, Ruey-Lung Hsiao, and D. Stott Parker. Map-
Reduce-Merge: Simplified relational data processing on large clusters. In Pro-
ceedings of the 2007 ACM SIGMOD International Conference on Management of
Data, pages 1029–1040, Beijing, China, 2007.
[37] Cheng-Tao Chu, Sang Kyun Kim, Yi-An Lin, YuanYuan Yu, Gary Bradski, An-
drew Ng, and Kunle Olukotun. Map-Reduce for machine learning on multicore.
In Advances in Neural Information Processing Systems 19 (NIPS 2006), pages
281–288, Vancouver, British Columbia, Canada, 2006.
[38] Kenneth W. Church and Patrick Hanks. Word association norms, mutual infor-
mation, and lexicography. Computational Linguistics, 16(1):22–29, 1990.
[39] Jonathan Cohen. Graph twiddling in a MapReduce world. Computing in Science
and Engineering, 11(4):29–41, 2009.
[40] Brian F. Cooper, Adam Silberstein, Erwin Tam, Raghu Ramakrishnan, and Rus-
sell Sears. Benchmarking cloud serving systems with YCSB. In Proceedings of the
First ACM Symposium on Cloud Computing (ACM SOCC 2010), Indianapolis,
Indiana, 2010.
7.3. MAPREDUCE AND BEYOND 161
[42] W. Bruce Croft, Donald Meztler, and Trevor Strohman. Search Engines: Infor-
mation Retrieval in Practice. Addison-Wesley, Reading, Massachusetts, 2009.
[43] David Culler, Richard Karp, David Patterson, Abhijit Sahay, Klaus Erik Schauser,
Eunice Santos, Ramesh Subramonian, and Thorsten von Eicken. LogP: Towards
a realistic model of parallel computation. ACM SIGPLAN Notices, 28(7):1–12,
1993.
[44] Doug Cutting, Julian Kupiec, Jan Pedersen, and Penelope Sibun. A practical
part-of-speech tagger. In Proceedings of the Third Conference on Applied Natural
Language Processing, pages 133–140, Trento, Italy, 1992.
[45] Jeffrey Dean and Sanjay Ghemawat. MapReduce: Simplified data processing on
large clusters. In Proceedings of the 6th Symposium on Operating System Design
and Implementation (OSDI 2004), pages 137–150, San Francisco, California, 2004.
[46] Jeffrey Dean and Sanjay Ghemawat. MapReduce: Simplified data processing on
large clusters. Communications of the ACM, 51(1):107–113, 2008.
[47] Jeffrey Dean and Sanjay Ghemawat. MapReduce: A flexible data processing tool.
Communications of the ACM, 53(1):72–77, 2010.
[49] Arthur P. Dempster, Nan M. Laird, and Donald B. Rubin. Maximum likelihood
from incomplete data via the EM algorithm. Journal of the Royal Statistical
Society. Series B (Methodological), 39(1):1–38, 1977.
[50] David J. DeWitt and Jim Gray. Parallel database systems: The future of high
performance database systems. Communications of the ACM, 35(6):85–98, 1992.
[51] David J. DeWitt, Randy H. Katz, Frank Olken, Leonard D. Shapiro, Michael R.
Stonebraker, and David Wood. Implementation techniques for main memory
database systems. ACM SIGMOD Record, 14(2):1–8, 1984.
[52] Mark Dredze, Alex Kulesza, and Koby Crammer. Multi-domain learning by
confidence-weighted parameter combination. Machine Learning, 79:123–149, 2010.
162 CHAPTER 7. CLOSING REMARKS
[53] Susan Dumais, Michele Banko, Eric Brill, Jimmy Lin, and Andrew Ng. Web
question answering: Is more always better? In Proceedings of the 25th Annual
International ACM SIGIR Conference on Research and Development in Informa-
tion Retrieval (SIGIR 2002), pages 291–298, Tampere, Finland, 2002.
[54] Chris Dyer, Aaron Cordova, Alex Mont, and Jimmy Lin. Fast, easy, and cheap:
Construction of statistical machine translation models with MapReduce. In Pro-
ceedings of the Third Workshop on Statistical Machine Translation at ACL 2008,
pages 199–207, Columbus, Ohio, 2008.
[56] Qin Gao and Stephan Vogel. Training phrase-based machine translation models on
the cloud: Open source machine translation toolkit Chaski. The Prague Bulletin
of Mathematical Linguistics, 93:37–46, 2010.
[57] Sanjay Ghemawat, Howard Gobioff, and Shun-Tak Leung. The Google File Sys-
tem. In Proceedings of the 19th ACM Symposium on Operating Systems Principles
(SOSP 2003), pages 29–43, Bolton Landing, New York, 2003.
[58] Seth Gilbert and Nancy Lynch. Brewer’s Conjecture and the feasibility of consis-
tent, available, partition-tolerant web services. ACM SIGACT News, 33(2):51–59,
2002.
[59] Michelle Girvan and Mark E. J. Newman. Community structure in social and
biological networks. Proceedings of the National Academy of Science, 99(12):7821–
7826, 2002.
[60] Ananth Grama, Anshul Gupta, George Karypis, and Vipin Kumar. Introduction
to Parallel Computing. Addison-Wesley, Reading, Massachusetts, 2003.
[61] Mark S. Granovetter. The strength of weak ties. The American Journal of Soci-
ology, 78(6):1360–1380, 1973.
[62] Mark S. Granovetter. The strength of weak ties: A network theory revisited.
Sociological Theory, 1:201–233, 1983.
[63] Zoltán Gyöngyi and Hector Garcia-Molina. Web spam taxonomy. In Proceedings
of the First International Workshop on Adversarial Information Retrieval on the
Web (AIRWeb 2005), pages 39–47, Chiba, Japan, 2005.
7.3. MAPREDUCE AND BEYOND 163
[64] Per Hage and Frank Harary. Island Networks: Communication, Kinship, and
Classification Structures in Oceania. Cambridge University Press, Cambridge,
England, 1996.
[65] Alon Halevy, Peter Norvig, and Fernando Pereira. The unreasonable effectiveness
of data. Communications of the ACM, 24(2):8–12, 2009.
[66] James Hamilton. On designing and deploying Internet-scale services. In Proceed-
ings of the 21st Large Installation System Administration Conference (LISA ’07),
pages 233–244, Dallas, Texas, 2007.
[67] James Hamilton. Cooperative Expendable Micro-Slice Servers (CEMS): Low cost,
low power servers for Internet-scale services. In Proceedings of the Fourth Bien-
nial Conference on Innovative Data Systems Research (CIDR 2009), Asilomar,
California, 2009.
[68] Jeff Hammerbacher. Information platforms and the rise of the data scientist.
In Toby Segaran and Jeff Hammerbacher, editors, Beautiful Data, pages 73–84.
O’Reilly, Sebastopol, California, 2009.
[69] Zelig S. Harris. Mathematical Structures of Language. Wiley, New York, 1968.
[70] Md. Rafiul Hassan and Baikunth Nath. Stock market forecasting using hidden
Markov models: A new approach. In Proceedings of the 5th International Confer-
ence on Intelligent Systems Design and Applications (ISDA ’05), pages 192–196,
Wroclaw, Poland, 2005.
[71] Bingsheng He, Wenbin Fang, Qiong Luo, Naga K. Govindaraju, and Tuyong
Wang. Mars: A MapReduce framework on graphics processors. In Proceedings
of the 17th International Conference on Parallel Architectures and Compilation
Techniques (PACT 2008), pages 260–269, Toronto, Ontario, Canada, 2008.
[72] Tony Hey, Stewart Tansley, and Kristin Tolle. The Fourth Paradigm: Data-
Intensive Scientific Discovery. Microsoft Research, Redmond, Washington, 2009.
[73] Tony Hey, Stewart Tansley, and Kristin Tolle. Jim Gray on eScience: A trans-
formed scientific method. In Tony Hey, Stewart Tansley, and Kristin Tolle, editors,
The Fourth Paradigm: Data-Intensive Scientific Discovery. Microsoft Research,
Redmond, Washington, 2009.
[74] John Howard, Michael Kazar, Sherri Menees, David Nichols, Mahadev Satya-
narayanan, Robert Sidebotham, and Michael West. Scale and performance in
a distributed file system. ACM Transactions on Computer Systems, 6(1):51–81,
1988.
164 CHAPTER 7. CLOSING REMARKS
[75] Michael Isard, Mihai Budiu, Yuan Yu, Andrew Birrell, and Dennis Fetterly. Dryad:
Distributed data-parallel programs from sequential building blocks. In Proceedings
of the ACM SIGOPS/EuroSys European Conference on Computer Systems 2007
(EuroSys 2007), pages 59–72, Lisbon, Portugal, 2007.
[76] Adam Jacobs. The pathologies of big data. ACM Queue, 7(6), 2009.
[77] Joseph JaJa. An Introduction to Parallel Algorithms. Addison-Wesley, Reading,
Massachusetts, 1992.
[78] Frederick Jelinek. Statistical methods for speech recognition. MIT Press, Cam-
bridge, Massachusetts, 1997.
[79] Daniel Jurafsky and James H. Martin. Speech and Language Processing. Pearson,
Upper Saddle River, New Jersey, 2009.
[80] U Kang, Charalampos Tsourakakis, Ana Paula Appel, Christos Faloutsos, and
Jure Leskovec. HADI: Fast diameter estimation and mining in massive graphs
with Hadoop. Technical Report CMU-ML-08-117, School of Computer Science,
Carnegie Mellon University, 2008.
[81] U Kang, Charalampos E. Tsourakakis, and Christos Faloutsos. PEGASUS: A
peta-scale graph mining system—implementation and observations. In Proceed-
ings of the 2009 Ninth IEEE International Conference on Data Mining (ICDM
2009), pages 229–238, Miami, Floria, 2009.
[82] Howard Karloff, Siddharth Suri, and Sergei Vassilvitskii. A model of computation
for MapReduce. In Proceedings of the 21st Annual ACM-SIAM Symposium on
Discrete Algorithms (SODA 2010), Austin, Texas, 2010.
[83] Aaron Kimball, Sierra Michels-Slettvet, and Christophe Bisciglia. Cluster com-
puting for Web-scale data processing. In Proceedings of the 39th ACM Techni-
cal Symposium on Computer Science Education (SIGCSE 2008), pages 116–120,
Portland, Oregon, 2008.
[84] Jon M. Kleinberg. Authoritative sources in a hyperlinked environment. Journal
of the ACM, 46(5):604–632, 1999.
[85] Philipp Koehn. Statistical Machine Translation. Cambridge University Press,
Cambridge, England, 2010.
[86] Philipp Koehn, Franz J. Och, and Daniel Marcu. Statistical phrase-based trans-
lation. In Proceedings of the 2003 Human Language Technology Conference of
the North American Chapter of the Association for Computational Linguistics
(HLT/NAACL 2003), pages 48–54, Edmonton, Alberta, Canada, 2003.
7.3. MAPREDUCE AND BEYOND 165
[87] John D. Lafferty, Andrew McCallum, and Fernando Pereira. Conditional random
fields: Probabilistic models for segmenting and labeling sequence data. In Pro-
ceedings of the Eighteenth International Conference on Machine Learning (ICML
’01), pages 282–289, San Francisco, California, 2001.
[88] Ronny Lempel and Shlomo Moran. SALSA: The Stochastic Approach for Link-
Structure Analysis. ACM Transactions on Information Systems, 19(2):131–160,
2001.
[89] Abby Levenberg, Chris Callison-Burch, and Miles Osborne. Stream-based transla-
tion models for statistical machine translation. In Proceedings of the 11th Annual
Conference of the North American Chapter of the Association for Computational
Linguistics (NAACL HLT 2010), Los Angeles, California, 2010.
[90] Abby Levenberg and Miles Osborne. Stream-based randomised language models
for SMT. In Proceedings of the 2009 Conference on Empirical Methods in Natural
Language Processing, pages 756–764, Singapore, 2009.
[91] Adam Leventhal. Triple-parity RAID and beyond. ACM Queue, 7(11), 2009.
[92] Jimmy Lin. An exploration of the principles underlying redundancy-based factoid
question answering. ACM Transactions on Information Systems, 27(2):1–55, 2007.
[93] Jimmy Lin. Exploring large-data issues in the curriculum: A case study with
MapReduce. In Proceedings of the Third Workshop on Issues in Teaching Com-
putational Linguistics (TeachCL-08) at ACL 2008, pages 54–61, Columbus, Ohio,
2008.
[94] Jimmy Lin. Scalable language processing algorithms for the masses: A case study
in computing word co-occurrence matrices with MapReduce. In Proceedings of the
2008 Conference on Empirical Methods in Natural Language Processing (EMNLP
2008), pages 419–428, Honolulu, Hawaii, 2008.
[95] Jimmy Lin, Anand Bahety, Shravya Konda, and Samantha Mahindrakar. Low-
latency, high-throughput access to static global resources within the Hadoop
framework. Technical Report HCIL-2009-01, University of Maryland, College
Park, Maryland, January 2009.
[96] Dong C. Liu, Jorge Nocedal, Dong C. Liu, and Jorge Nocedal. On the limited
memory BFGS method for large scale optimization. Mathematical Programming
B, 45(3):503–528, 1989.
[97] Adam Lopez. Statistical machine translation. ACM Computing Surveys, 40(3):1–
49, 2008.
166 CHAPTER 7. CLOSING REMARKS
[98] Grzegorz Malewicz, Matthew H. Austern, Aart J. C. Bik, James C. Dehnert, Ilan
Horn, Naty Leiser, and Grzegorz Czajkowski. Pregel: A system for large-scale
graph processing. In Proceedings of the 28th ACM Symposium on Principles of
Distributed Computing (PODC 2009), page 6, Calgary, Alberta, Canada, 2009.
[104] Michael D. McCool. Scalable programming models for massively multicore pro-
cessors. Proceedings of the IEEE, 96(5):816–831, 2008.
[105] Marshall K. McKusick and Sean Quinlan. GFS: Evolution on fast-forward. ACM
Queue, 7(7), 2009.
[106] John Mellor-Crummey, David Whalley, and Ken Kennedy. Improving memory
hierarchy performance for irregular applications using data and computation re-
orderings. International Journal of Parallel Programming, 29(3):217–247, 2001.
[107] Donald Metzler, Jasmine Novak, Hang Cui, and Srihari Reddy. Building enriched
document representations using aggregated anchor text. In Proceedings of the
32nd Annual International ACM SIGIR Conference on Research and Development
in Information Retrieval (SIGIR 2009), pages 219–226, 2009.
[108] David R. H. Miller, Tim Leek, and Richard M. Schwartz. A hidden Markov model
information retrieval system. In Proceedings of the 22nd Annual International
ACM SIGIR Conference on Research and Development in Information Retrieval
(SIGIR 1999), pages 214–221, Berkeley, California, 1999.
7.3. MAPREDUCE AND BEYOND 167
[109] Alistair Moffat, William Webber, and Justin Zobel. Load balancing for term-
distributed parallel retrieval. In Proceedings of the 29th Annual International
ACM SIGIR Conference on Research and Development in Information Retrieval
(SIGIR 2006), pages 348–355, Seattle, Washington, 2006.
[110] Kamal Nigam, John Lafferty, and Andrew McCallum. Using maximum entropy
for text classification. In Proceedings of the IJCAI-99 Workshop on Machine
Learning for Information Filtering, pages 61–67, Stockholm, Sweden, 1999.
[111] Daniel Nurmi, Rich Wolski, Chris Grzegorczyk, Graziano Obertelli, Sunil Soman,
Lamia Youseff, and Dmitrii Zagorodnov. The Eucalyptus open-source cloud-
computing system. In Proceedings of the 9th IEEE/ACM International Sym-
posium on Cluster Computing and the Grid, pages 124–131, Washington, D.C.,
2009.
[112] Franz J. Och and Hermann Ney. A systematic comparison of various statistical
alignment models. Computational Linguistics, 29(1):19–51, 2003.
[113] Christopher Olston and Marc Najork. Web crawling. Foundations and Trends in
Information Retrieval, 4(3):175–246, 2010.
[114] Christopher Olston, Benjamin Reed, Utkarsh Srivastava, Ravi Kumar, and An-
drew Tomkins. Pig Latin: A not-so-foreign language for data processing. In Pro-
ceedings of the 2008 ACM SIGMOD International Conference on Management of
Data, pages 1099–1110, Vancouver, British Columbia, Canada, 2008.
[115] Kunle Olukotun and Lance Hammond. The future of microprocessors. ACM
Queue, 3(7):27–34, 2005.
[116] Sean Owen and Robin Anil. Mahout in Action. Manning Publications Co., Green-
wich, Connecticut, 2010.
[117] Lawrence Page, Sergey Brin, Rajeev Motwani, and Terry Winograd. The Page-
Rank citation ranking: Bringing order to the Web. Stanford Digital Library Work-
ing Paper SIDL-WP-1999-0120, Stanford University, 1999.
[118] Bo Pang and Lillian Lee. Opinion mining and sentiment analysis. Foundations
and Trends in Information Retrieval, 2(1–2):1–135, 2008.
[119] David A. Patterson. The data center is the computer. Communications of the
ACM, 52(1):105, 2008.
[120] Andrew Pavlo, Erik Paulson, Alexander Rasin, Daniel J. Abadi, David J. De-
Witt, Samuel Madden, and Michael Stonebraker. A comparison of approaches to
168 CHAPTER 7. CLOSING REMARKS
[121] Sasa Petrovic, Miles Osborne, and Victor Lavrenko. Streaming first story detec-
tion with application to Twitter. In Proceedings of the 11th Annual Conference
of the North American Chapter of the Association for Computational Linguistics
(NAACL HLT 2010), Los Angeles, California, 2010.
[122] Rob Pike, Sean Dorward, Robert Griesemer, and Sean Quinlan. Interpreting the
data: Parallel analysis with Sawzall. Scientific Programming Journal, 13(4):277–
298, 2005.
[123] Eduardo Pinheiro, Wolf-Dietrich Weber, and Luiz André Barroso. Failure trends
in a large disk drive population. In Proceedings of the 5th USENIX Conference
on File and Storage Technologies (FAST 2007), San Jose, California, 2008.
[124] Xiaoguang Qi and Brian D. Davison. Web page classification: Features and algo-
rithms. ACM Computing Surveys, 41(2), 2009.
[125] Lawrence R. Rabiner. A tutorial on hidden Markov models and selected appli-
cations in speech recognition. In Readings in Speech Recognition, pages 267–296.
Morgan Kaufmann Publishers, San Francisco, California, 1990.
[126] M. Mustafa Rafique, Benjamin Rose, Ali R. Butt, and Dimitrios S. Nikolopou-
los. Supporting MapReduce on large-scale asymmetric multi-core clusters. ACM
Operating Systems Review, 43(2):25–34, 2009.
[127] Colby Ranger, Ramanan Raghuraman, Arun Penmetsa, Gary Bradski, and Chris-
tos Kozyrakis. Evaluating MapReduce for multi-core and multiprocessor systems.
In Proceedings of the 13th International Symposium on High-Performance Com-
puter Architecture (HPCA 2007), pages 205–218, Phoenix, Arizona, 2007.
[128] Delip Rao and David Yarowsky. Ranking and semi-supervised classification
on large scale graphs using Map-Reduce. In Proceedings of the ACL/IJCNLP
2009 Workshop on Graph-Based Methods for Natural Language Processing
(TextGraphs-4), Singapore, 2009.
[129] Michael A. Rappa. The utility business model and the future of computing ser-
vices. IBM Systems Journal, 34(1):32–42, 2004.
[131] Thomas Sandholm and Kevin Lai. MapReduce optimization using regulated dy-
namic prioritization. In Proceedings of the Eleventh International Joint Confer-
ence on Measurement and Modeling of Computer Systems (SIGMETRICS ’09),
pages 299–310, Seattle, Washington, 2009.
[132] Michael Schatz. High Performance Computing for DNA Sequence Alignment and
Assembly. PhD thesis, University of Maryland, College Park, 2010.
[133] Frank Schmuck and Roger Haskin. GPFS: A shared-disk file system for large
computing clusters. In Proceedings of the First USENIX Conference on File and
Storage Technologies, pages 231–244, Monterey, California, 2002.
[134] Donovan A. Schneider and David J. DeWitt. A performance evaluation of four
parallel join algorithms in a shared-nothing multiprocessor environment. In Pro-
ceedings of the 1989 ACM SIGMOD International Conference on Management of
Data, pages 110–121, Portland, Oregon, 1989.
[135] Bianca Schroeder, Eduardo Pinheiro, and Wolf-Dietrich Weber. DRAM errors
in the wild: A large-scale field study. In Proceedings of the Eleventh Interna-
tional Joint Conference on Measurement and Modeling of Computer Systems
(SIGMETRICS ’09), pages 193–204, Seattle, Washington, 2009.
[136] Hinrich Schütze. Automatic word sense discrimination. Computational Linguis-
tics, 24(1):97–123, 1998.
[137] Hinrich Schütze and Jan O. Pedersen. A cooccurrence-based thesaurus and two
applications to information retrieval. Information Processing and Management,
33(3):307–318, 1998.
[138] Satoshi Sekine and Elisabete Ranchhod. Named Entities: Recognition, Classifica-
tion and Use. John Benjamins, Amsterdam, The Netherlands, 2009.
[139] Kristie Seymore, Andrew Mccallum, and Ronald Rosenfeld. Learning hidden
Markov model structure for information extraction. In Proceedings of the AAAI-99
Workshop on Machine Learning for Information Extraction, pages 37–42, Or-
lando, Florida, 1999.
[140] Fei Sha and Fernando Pereira. Shallow parsing with conditional random
fields. In Proceedings of the 2003 Human Language Technology Conference of
the North American Chapter of the Association for Computational Linguistics
(HLT/NAACL 2003), pages 134–141, Edmonton, Alberta, Canada, 2003.
[141] Noah Smith. Log-linear models. http://www.cs.cmu.edu/~nasmith/papers/
smith.tut04.pdf, 2004.
170 CHAPTER 7. CLOSING REMARKS
[142] Christopher Southan and Graham Cameron. Beyond the tsunami: Developing the
infrastructure to deal with life sciences data. In Tony Hey, Stewart Tansley, and
Kristin Tolle, editors, The Fourth Paradigm: Data-Intensive Scientific Discovery.
Microsoft Research, Redmond, Washington, 2009.
[143] Mario Stanke and Stephan Waack. Gene prediction with a hidden Markov model
and a new intron submodel. Bioinformatics, 19 Suppl 2:ii215–225, October 2003.
[144] Michael Stonebraker, Daniel Abadi, David J. DeWitt, Sam Madden, Erik Paulson,
Andrew Pavlo, and Alexander Rasin. MapReduce and parallel DBMSs: Friends
or foes? Communications of the ACM, 53(1):64–71, 2010.
[145] Alexander S. Szalay, Peter Z. Kunszt, Ani Thakar, Jim Gray, Don Slutz, and
Robert J. Brunner. Designing and mining multi-terabyte astronomy archives:
The Sloan Digital Sky Survey. SIGMOD Record, 29(2):451–462, 2000.
[146] Wittawat Tantisiriroj, Swapnil Patil, and Garth Gibson. Data-intensive file sys-
tems for Internet services: A rose by any other name. . . . Technical Report CMU-
PDL-08-114, Parallel Data Laboratory, Carnegie Mellon University, 2008.
[147] Chandramohan A. Thekkath, Timothy Mann, and Edward K. Lee. Frangipani: A
scalable distributed file system. In Proceedings of the 16th ACM Symposium on
Operating Systems Principles (SOSP 1997), pages 224–237, Saint-Malo, France,
1997.
[148] Leslie G. Valiant. A bridging model for parallel computation. Communications
of the ACM, 33(8):103–111, 1990.
[149] Luis M. Vaquero, Luis Rodero-Merino, Juan Caceres, and Maik Lindner. A break
in the clouds: Towards a cloud definition. ACM SIGCOMM Computer Commu-
nication Review, 39(1):50–55, 2009.
[150] Stephan Vogel, Hermann Ney, and Christoph Tillmann. HMM-based word align-
ment in statistical translation. In Proceedings of the 16th International Confer-
ence on Computational Linguistics (COLING 1996), pages 836–841, Copenhagen,
Denmark, 1996.
[151] Yi Wang, Hongjie Bai, Matt Stanton, Wen-Yen Chen, and Edward Y. Chang.
PLDA: Parallel latent Dirichlet allocation for large-scale applications. In Proceed-
ings of the Fifth International Conference on Algorithmic Aspects in Information
and Management (AAIM 2009), pages 301–314, San Francisco, California, 2009.
[152] Duncan J. Watts and Steven H. Strogatz. Collective dynamics of ‘small-world’
networks. Nature, 393:440–442, 1998.
7.3. MAPREDUCE AND BEYOND 171