AI-Systems

Machine Learning
in the Cloud
Joseph E. Gonzalez
Co-director of the RISE Lab
[email protected]
What is cloud computing?
“The interesting thing about Cloud
Computing is that we’ve redefined
Cloud Computing to include
everything that we already do. . . . I
don’t understand what we would do
differently in the light of Cloud
Computing other than change the
wording of some of our ads.”
-- Larry Ellison,
Wall Street Journal, 2008

3
Quote from “Above the Clouds: A Berkeley View of Cloud Computing”
8 years later …

2016
4
“If ‘cloud computing’ has a meaning,
it is not a way of doing computing, but
rather a way of thinking about
computing: a devil-may-care
approach which says, ‘Don't ask
questions. Don't worry about who
controls your computing or who holds
your data. Don't check for a hook
hidden inside our service before you
swallow it. Trust companies without
hesitation.’ In other words, ‘Be a
sucker.’ ”
-- Richard Stallman,
Boston Review, 2010
5
https://bostonreview.net/articles/richard-stallman-free-software-drm/
2006-2011

Early paper on cloud Published in 2009 and received over

computing and helped drive 8,4000 citations

excitement in the field.

Made the case for cloud computing

 Illusion of infinite resources
 Elimination of up-front costs
 Pay-per-use
 Economies of scale for everyone
 Simplified operations
 Higher hardware utilization
 Defining Characteristics of the Cloud
 The illusion of infinite computing resources available on demand
 The elimination of an up-front commitment by users
 The ability to pay for use of computing resources on a short-term basis as needed

 Essentially Utility Computing

Public Cloud: when these services are available to the general public
Private Cloud: when these services are sold within a business

Quote from “Above the Clouds: A Berkeley View of Cloud Computing”

The Beginning of the Cloud
1961: John McCarthy presents idea of Utility Computing
2006: Amazon Web Services (AWS) Launched
2008: Google Cloud Launched
2008: Azure Cloud Launched
2009: Berkeley writes: “Above the Clouds: A Berkeley View of Cloud
Computing”
Utility Computing is an old idea
“If computers of the kind I have advocated become the computers of the
future, then computing may someday be organized as a public utility just
as the telephone system is a public utility... The computer utility could
become the basis of a new and important industry.”
– John McCarthy
Public Cloud
(>$300B)

Early Pioneer in Artificial Intelligence

Turing award for contributions to AI
9
Why did it take until 2006?
 Change is business models for computing
 Web2.0 – low commitment, self-serve, pay-as-you-go -- not enough capital
to own equipment.

 Existing big-tech companies were aggressively exploiting economies

of scale to drive down costs
 Renting becomes cheaper than buying machines

 New workloads emerged to leverage elasticity:

 Web applications needed to handle demand surges
 Data intensive apps leveraged scale -- fast is cheap

10
Economics of the Cloud
 CapEx to OpEx: transition from large up-front capital expenditures
to operational expenditures
 More money to spend on your launching your business

 Improved Utilization through statistical multiplexing

 real-world server utilization for a single business ~ 5% to 20%.

 Economies of scales
 Negotiate lower hardware prices
 Spread management costs
 Leverage existing investments

11
The Cloud Enabled Academic Research
 Access to the latest hardware
 Ability to burst experiments near conference deadlines
 Usually…

 Ability for students to build and evaluate large-scale systems

 I would frequently run concurrent experiments with hundreds of machines
each!

 industrial adoption
 Companies can evaluate open-source (academic) big data tools without big
upfront investment in hardware.
What about the Cloud?
 Access to latest GPUs and TPUs drove AI research
 used a LOT OF CREDITS (thank you AWS, Azure, & Google!)

Conference
Deadlines

Y-axis scales are hidden We only had to “pay” for

because they are this region.
embarrassing

13
The Elasticity
of the cloud
drove us to rethink
our approach to AI Research

14
 Deadline

Accuracy

Time
 Deadline
Accuracy

Time
 Deadline

Accuracy

Time
 Exploration Exploitation
00:00 Time (mins) 60:00 00:00 Time (mins) 45:00

GPU 1 GPU 1
GPU 2 GPU 2
GPU 3 GPU 3
Model GPU 4 GPU 4
learning GPU 5
GPU 5
rate 𝒍, weight
decay d GPU 6 GPU 6
GPU 7 GPU 7
GPU 8 GPU 8
 Linear
Exploitation Scaling
00:00 Time (mins) 45:00
GPU 1
GPU 2
GPU 3
GPU 4
GPU 5
GPU 6
GPU 7
GPU 8
 Linear
Exploitation Scaling
00:00 Time (mins) 45:00
GPU 1
GPU 2
GPU 3
GPU 4
GPU 5
GPU 6
GPU 7
GPU 8
Fixed Cluster Cloud
Resources = Machines Resources = Money
Fixed Cluster Cloud
Resources = Machines Resources = Money

GPU-mins
GPU-mins

start Time deadline

Fixed Cluster Cloud
Resources = Machines Resources = Money

More parallel exploration

GPU-mins
Less Parallel
Exploitation
GPU-mins

start Time deadline

start Time deadline

 00:00 Time (mins) Deadline

GPU 1
GPU 2
GPU 3
GPU 4

Liaw et al.
 00:00 Time (mins) Deadline

GPU 1
GPU 2
GPU 3
GPU 4
GPU 5
 00:00 Time (mins) Deadline

GPU 1
GPU 2
GPU 3
GPU 4
GPU 5
 00:00 Time (mins) Deadline

GPU 1
GPU 2
GPU 3
GPU 4
GPU 5
GPU 6
 00:00 Time (mins) Deadline

GPU 1
GPU 2
GPU 3
GPU 4
GPU 5
GPU 6
 00:00 Time (mins) Deadline

GPU 1
GPU 2
GPU 3
GPU 4
GPU 5
GPU 6
 When it comes to machine-time allocation
00:00 Time (mins) Deadline 00:00 Time (mins) Deadline

GPU 1 GPU 1

GPU 2 GPU 2

GPU 3 GPU 3

GPU 4 GPU 4
GPU 5
GPU 6

Triangles > Rectangles

Think outside the box
Compute

Cloud
Cluster Computing

Compute
Computing
e
e
m

Tim
Ti

Storage Storage

Cloud computing  Infinite resources with a finite budget (volume

constraints). 31
 Cloud
Computing
Compute

e
Tim
Storage

The cloud as a utility

“Serverless”

32
33
©2019 RISELab
Canonical example
Canonical example

BaaS

FaaS

BaaS
BaaS
 AWS Lambda

Introduced Function as a Service (FaaS)

Autoscaling done right
Highly elastic - adapts quickly
Scales down to zero
Fine-grained 100 ms billing increment
Cloud provider shares risk and responsibility for utilization

Strong isolation allowing multi-tenant multiplexing

Benefits from scale of Amazon’s platform & ecosystem of APIs

36
©2019 RISELab
A Layer to Simplify Using the Cloud
Three essential qualities of serverless
computing
 H
‌ ides servers and the complexity of programming and operating
them
 Offers a pay-per-use cost model with no charge for idle resources
 Has excellent autoscaling so resources match demand closely
Airport Analogy
When you arrive at the destination airport and need to get to your hotel
you could:
1. Buy a car and drive [Legacy on premise systems]
 Long term investment and you are responsible for everything

2. Rent a car and drive [Serverfull]

 You are still responsible for fuel, parking, insurance, …

3. Take an Uber. [Serverless]

 You are paying only for the transportation you need
Where is the cloud headed?
With each phase of the cloud
(and computing), we raised the
abstraction.

This continuous process of

simplification enables
concurrent innovation on
both sides of the boundary.

It also improves portability and

transforms the market.
Sky Computing
Sky Example ML Pipeline
Data
Training Serving
proc

Sky

Secure
Training Serving
Data proc

● Use Azure Confidential Computing for secure data processing

● Use Google Cloud for training on TPUs (fastest and cheapest)
● Use AWS for serving on Inferentia (cheapest) 45
Conjectures for the Future
● We will continue to race towards utility-oriented computing
○ Pay for consumption and not capacity

● Higher levels of abstraction will

○ Reduce operational complexity (burden shifts to the cloud + ML)
○ Drive more rapid innovation in cloud hardware
○ Enable applications to more easily span multiple clouds

● Sky computing is the inevitable future of computing

46
Readings This Week

47
Pollux: Co-adaptive Cluster Scheduling for
Goodput-Optimized Deep Learning
 Published in OSDI’21 – Best Paper Award
 Why we chose it?
 Good example of work exploring scheduling of for ML
 Addresses ML and Systems concerns: throughput, improvements in
accuracy, fairness?

 Things to think about:

 Implications for elasticity?
 What about hyperparameter search?

48
The Sky Above the Clouds [Unpublished]
 Draft of the vision paper describing Sky research agenda
 Do Not Distribute
 Feedback will help the paper (be critical!)

 Makes a case for both the inevitability and need for research in
“Sky Computing”
 Things to think about:
 Presentation of premise [what is proposed?]
 Role of data
 Role of research
 ML Systems Research Case?
49
FrugalML: How to Use ML Prediction APIs
More Accurately and Cheaply
 Published in NuerIPS’20
 Example of a “Sky Computing” ML research direction
 Combining competing prediction services to improve accuracy and reduce
costs.
 Potentially exciting new research direction!

 Things to think about:

 Latency
 Out-of-domain performance
 Uncertainty calibration and model biases
 Mathematical presentations
50