See

discussions, stats, and author profiles for this publication at: http://www.researchgate.net/publication/2830193

Simulation of Dynamic Grid Replication

Strategies in OptorSim
ARTICLE AUGUST 2003
Source: CiteSeer

CITATIONS

DOWNLOADS

VIEWS

74

183

134

5 AUTHORS, INCLUDING:
William Hamish Bell

A. Paul Millar

University of Geneva

Deutsches Elektronen-Synchrotron

137 PUBLICATIONS 2,125 CITATIONS

38 PUBLICATIONS 1,021 CITATIONS

SEE PROFILE

SEE PROFILE

Kurt Stockinger
Zurich University of Applied Sciences
115 PUBLICATIONS 3,026 CITATIONS
SEE PROFILE

Available from: Kurt Stockinger

Retrieved on: 20 September 2015

CENTRO PER LA RICERCA

SCIENTIFICA E TECNOLOGICA

38050 Povo (Trento), Italy

Tel.: +39 0461 314312
Fax: +39 0461 302040
email: [email protected] url: http://www.itc.it

SIMULATION OF DYNAMIC GRID REPLICATION

STRATEGIES IN OPTORSIM
Bell W.B., Cameron D.G., Capozza L.,
Millar A.P., Stockinger K., Zini F.
July 2002
Technical Report # 020719

Istituto Trentino di Cultura, 2002

LIMITED DISTRIBUTION NOTICE

This report has been submitted forpublication outside of ITC and will probably be copyrighted if accepted for publication. It has been
issued as a Technical Report forearly dissemination of its contents. In view of the transfert of copy right tothe outside publisher, its
distribution outside of ITC priorto publication should be limited to peer communications and specificrequests. After outside publication,
material will be available only inthe form authorized by the copyright owner.

Simulation of Dynami Grid Repli ation

Strategies in OptorSim
William H. Bell1 , David G. Cameron1 , Luigi Capozza2 ,
A. Paul Millar1 , Kurt Sto kinger3 , Floriano Zini2
1

University of Glasgow, Glasgow, G12 8QQ, S otland

ITC-irst, Via Sommarive 18, 38050 Povo (Trento), Italy
CERN, European Organization for Nu lear Resear h, 1211 Geneva, Switzerland
2

Abstra t. Computational Grids normally deal with large omputationally intensive problems on small data sets. In ontrast, Data Grids mostly
deal with large omputational problems that in turn require evaluating
and mining large amounts of data. Repli ation is regarded as one of the
major optimisation te hniques for providing fast data a ess.
Within this paper, several repli ation algorithms are studied. This is
a hieved using the Grid simulator: OptorSim. OptorSim provides a modular framework within whi h optimisation strategies an be studied under dierent Grid ongurations. The goal is to explore the stability and
transient behaviour of sele ted optimisation te hniques.

1 Introdu tion
Within the Grid ommunity mu h work has been done on providing the basi infrastru ture for a typi al Grid environment. Globus [4, Condor [1 and re ently
the EU DataGrid [3 have ontributed substantially to ore Grid middleware
servi es software that are available as the basis for further appli ation development. However, little eort has been made so far to optimise the use of Grid
resour es.
To use a Data Grid, users typi ally submit job s. In order for a job to be
exe uted, three types of resour es are required: omputing fa ilities, data a ess
and storage, and network onne tivity. The Grid must make s heduling de isions
for ea h job based on the urrent state of these resour es (workload and features
of Computing Elements, lo ation of data, network load). Complete optimisation
is a hieved when the ombined resour e impa t of all jobs is minimised, allowing
jobs to run as fast as possible.
File repli ation (i.e. spread of multiple opies of les a ross the Grid) is an
ee tive te hnique for redu ing data a ess overhead. Maintaining an optimal
distribution of repli as implies that the Grid optimisation servi e [7 must be able
to modify the geographi lo ation of data les. This is a hieved by triggering
both repli ation and deletion of data les. By re e ting the dynami load on
the Grid, su h repli a management will ae t the migration of parti ular les
toward sites that show in reased frequen y of le-a ess requests.

In order to study the omplex nature of a typi al Grid environment and evaluate various repli a optimisation algorithms, a Grid simulator ( alled OptorSim)
was developed. In this paper the design on epts of OptorSim are dis ussed and
preliminary results based on sele ted repli ation algorithms are reported.
The paper is stru tured as follows. Se tion 2 des ribes the design of the
simulator OptorSim. Various repli ation algorithms are dis ussed in Se tion 3.
After setting the simulation onguration in Se tion 4, Se tion 5 is dedi ated
to a des ription of simulation results. Se tion 6 highlights related work. Finally,
Se tion 7 on ludes the paper and reports on future work.

2 Simulation Design
OptorSim [2 is a simulation pa kage written in Java. It was developed to study

the ee tiveness of repli a optimisation algorithms within a Data Grid environment.

2.1 Ar hite ture

One of the main design onsiderations for OptorSim is to model the intera tions
of the individual Grid omponents of a running Data Grid as realisti ally as possible. Therefore, the simulation is based on the ar hite ture of the EU DataGrid
proje t [14 as illustrated in Figure 1.
User Interface

Resource Broker

Job execution site

Job execution site

Replica

Manager

Replica
Optimiser

Computing
Element

Computing
Element

Storage
Element

Fig. 1.

Job execution site

Replica Manager

Replica Manager

Replica
Optimiser

Replica
Optimiser

Storage
Element

Computing
Element

Storage
Element

Simulated DataGrid Ar hite ture.

The simulation was onstru ted assuming that the Grid onsists of several
sites, ea h of whi h may provide omputational and data-storage resour es for
submitted jobs. Ea h site onsists of zero or more Computing Elements and zero

or more Storage Elements. Computing Elements run jobs, whi h use the data in
les stored on Storage Elements and a Resour e Broker ontrols the s heduling
of jobs to Computing Elements. Sites without Storage or Computing Elements
a t as network nodes or routers.
The de ision about data movement asso iated with jobs between sites is performed by a omponent alled the Repli a Manager. Within the Repli a Manager
the de ision to reate or delete repli as is ontrolled by a Repli a Optimiser alled
Optor. At the heart of Optor is a repli a optimisation algorithm, the properties
of whi h are dis ussed in Se tion 3.

2.2 Internals
In the simulation ea h Computing Element is represented by a thread. Job submission to the Computing Elements is managed by another thread: the Resour e
Broker. The Resour e Broker ensures every Computing Element ontinuously
runs jobs by frequently attempting to distribute jobs to all the Computing Elements. When the Resour e Broker nds an idle Computing Element, it sele ts
a job to run on it a ording to the poli y of the Computing Element, i.e. whi h
type of jobs it will run and how often it will run ea h job.
At any time, a Computing Element will be running at most one job. As soon
as the job nishes, another is assigned by the Resour e Broker. So, although there
is no expli it job s heduling algorithm, all Computing Elements pro ess jobs for
the duration of the simulation but are never overloaded. Currently, optimisation
only o urs after a job has been s heduled to a Computing Element. The more
omplex s enario of optimising both job s heduling and data a ess will be part
of future work.
Ea h job has a set of les it may request. Two types of referen e may be used
for a le: a logi al le name (LFN) and a physi al le name (PFN). An LFN is
an abstra t referen e to a le that is independent of both where the le is stored
and how many repli as exist. A PFN refers to a spe i repli a of some LFN,
lo ated at a denite site. Ea h LFN will have one PFN for ea h repli a in the
Grid.
A job will typi ally request a set of LFNs for data a ess. The order in whi h
those les are requested is determined by the a ess pattern. The following a ess
patterns were onsidered: sequential (the set of LFNs is ordered, forming a list
of su essive requests), random (les are sele ted randomly from set with a at
probability distribution), unitary random walk (set is ordered and su essive le
requests are exa tly one element away from previous le request, dire tion is
random) and Gaussian random walk (as with unitary random walk, but les are
sele ted from a Gaussian distribution entred on the previous le request).
When a le is required by a job, the le's LFN is used to lo ate the best repli a
via the Repli a Optimiser fun tion getBestFile(LFN, destinationStorageElement), where destinationStorageElement is the Storage Element to whi h
the repli a may be opied. It is assumed the Computing Element on whi h the
job is running and requested Storage Element are lo ated at the same site.

getBestFile() he ks the Repli a Catalogue for opies of the le. The Repli a Catalogue is a Grid middleware servi e urrently implemented within the simulation as a table of LFNs and all orresponding PFNs. By examining the available bandwidth between destinationStorageElement and all sites on whi h
a repli a of the le is stored, getBestFile() an hoose the PFN that will be
a essed fastest and hen e de rease the job running time.
The simulated version of getBestFile() partially fulls the fun tionality
as des ribed in [7. It is a blo king all that may ause repli ation to a Storage Element lo ated in the site where the job is running. After any repli ation
has ompleted, the PFN of the best available repli a is returned to the job. If
repli ation has not o urred, the best repli a is lo ated on a remote site and is
a essed by the job using remote I/O.
Both the repli ation time (if repli ation o urs) and the le a ess time (if
from a remote site) are dependent on the network hara teristi s over the duration of the onne tion. At any time, the bandwidth available to a transfer is
limited by the lowest bandwidth along the transfer path. For transfers utilising
a ommon network element, the bandwidth of that element is shared so ea h
transfer re eives an equal share.

3 Optimisation Algorithms
Repli a optimisation algorithms are the ore of the Repli a Optimiser. Over
the duration of a submitted job, PFNs for ea h LFN are requested by alling
getBestFile(). Optimisation algorithms implement getBestFile() so that it
may opy the requested le from the remote site to a Storage Element on the
same site as the requesting Computing Element. If all Storage Elements on this
site are full then a le must be deleted for the repli ation to su eed.
The strategy used to de ide whi h le should be deleted dierentiates optimisation algorithms. In the following, we brie y present three simple algorithms
and a more sophisti ated one in greater detail. These algorithms have been implemented into OptorSim.

3.1 Simple Algorithms

No repli ation. This algorithm never repli ates a le. The distribution of initial
le repli as is de ided at the beginning of the simulation and does not hange
during its exe ution. This algorithm returns a PFN with the largest expe ted
bandwidth. Sin e the network load varies during the simulation, the optimal
PFN may hange.
Un onditional repli ation, oldest le deleted. This algorithm always repli ates a le to the site where the job is exe uting. If there is no spa e to a ommodate the repli ation, the oldest le in the Storage Element is deleted.

Un onditional repli ation, least a essed le deleted. This algorithms

behaves as the previous method, ex ept the least a essed le in the past time
interval t is deleted.

3.2 An E onomi Approa h

This se tion presents a repli ation strategy based on an e onomi model for Grid
resour e optimisation. A general des ription of this e onomi approa h an be
found in [9.
The e onomi model we propose in ludes a tors (autonomous goal-seeking
entities) and the resour es in the Grid. Optimisation is a hieved via intera tion of
the a tors in the model, whose goals are maximising the prots and minimising
the osts of data resour e management. Data les represent the goods in the
market. They are pur hased by Computing Elements for jobs and by Storage
Elements in order to make an investment that will improve their revenues in the
future. They are sold by Storage Elements to Computing Elements and to other
Storage Elements. Computing Elements try to minimise the le pur hase ost,
while Storage Elements have the goal of maximising prots.
This e onomi model is utilised to solve two distin t problems: in de iding
if repli ation should o ur and in the sele tion of the expendable le(s) when
reating spa e for a new repli a.
When a job running on a Computing Element requests a le, the optimisation
tries to lo ate the heapest opy of it in the Grid by starting an au tion. Storage
Elements that have the le lo ally may reply, bidding a pri e that indi ates the
le transfer ost. A site that does not have a le lo ally may initiate its own
au tion to establish if, by repli ation, it an satisfy the le request. This me hanisms realises the global optimisation mentioned above. Currently, the au tion
proto ol has still to be integrated into OptorSim. In the following dis ussion, we
used the simpler proto ol des ribed in Se tion 2.2.
The me hanism for de iding if repli ation should o ur is implemented in
OptorSim. It is des ribed the following se tion.

Repli ation De ision. Within our e onomi model the Repli a Optimiser

needs to make an informed de ision about whether it should repli ate a le

to a lo al Storage Element. This de ision is based on whether the repli ation
(with asso iated le transfer and le deletion) will result in redu ed expe ted
future le a ess ost for the lo al Computing Element.
In order to make this de ision, the Repli a Optimiser keeps tra k of the
le requests it re eives and uses this history as input to an evaluation fun tion
E (f; r; n). This fun tion, dened in [9, returns the predi ted number of times a
le f will be requested in the next n requests based on the past r requests in
the history.
After any new le request is re eived by the Repli a Optimiser (say, for le
f ), the predi tion fun tion E is al ulated for f and every le in the storage.
If there is no le in the Storage Element that has a value less than the value

of f then no repli ation o urs. Otherwise, the least valuable le is sele ted for
deletion and a new repli a of f is reated on the Storage Element. If multiple
les on the Storage Element share the minimum value, the le having the earliest
last a ess time is deleted.
The evaluation fun tion E (f; r; n) is dened by the equation
E (f; r; n)

n
X
i=1

pi (f );

(1)

with the following argument.

Assuming that requests for les ontaining similar data are lustered in spatial and time lo ality, the request history an be des ribed as a random walk
in the spa e of integer le identiers1 . In the random walk, the identier of the
next requested le is obtained from the urrent identier by the addition of a
step, the value of whi h is given by some probability distribution. Assuming a
binomial distribution of the steps, the probability of re eiving a request for le
f at step i of the random walk is given by the equation
pi (f )

2iS
22iS id(f ) s + iS
1

jid(f )

j  iS

s

(2)

where s is the mean value of the binomial distribution , S is the maximum value
for the step, and id(f ) is a unique le identier (for instan e, the LFN). Then,
the most probable number of times le f will be requested during the next n
requests is given by (1).
A time interval t des ribes how far ba k the history goes and thus determines the number r of previous requests whi h are onsidered in the predi tion
fun tion. We assume that the mean arrival rate of requests is onstant. On e t
has been de ided, n is obtained by
n

=r

t0
t

(3)

where t is the future interval for whi h we intend to do the predi tion.
The value for S in (2) depends on the value of r. The mean value s is obtained from the re ent values of the step in the random walk. In parti ular, s
is al ulated as the weighted average of the last r steps, where weights de rease
over past time.
0

4 Simulation Conguration

4.1 Grid Conguration
The study of optimisation algorithms was arried out using a model of EU
DataGrid TestBed 1 sites and their asso iated network geometry as illustrated
1

We assume a mapping between le names and identiers that preserve le ontent
similarity.

in Figure 2. Within this model, ea h site was allo ated storage resour es proportional to their a tual hardware allo ations. Ea h TestBed site, ex luding CERN,
was assigned a Computing and Storage Element. CERN was allo ated a Storage Element to hold all of the master les but was not assigned a Computing
Element. Routers, as previously stated, were des ribed by reating a site without Computing or Storage Elements. The size of the Storage Elements for ea h
TestBed site are given in Table 1.

0110

RAL

0110

00111100 155M 1010

NorduGrid

Imperial College

2.5G

11
00

2.5G

10G

0110
11001100

NIKHEF
2.5G

10G

00111100
11
00
00
11
Lyon

0110 Testbed site

11001100 Router

2.5G

10G

2.5G

155M

10G
155M

622M

0110
CERN

0011

00111100
10G

Padova
1010
1
0
Milano 1
0 10M 0011 45M 10
Bologna
1010 45M 155M 10
10G

1G

Torino

10M

0110

Catania

The EU DataGrid TestBed 1 sites and the approximate network geometry. The
numbers indi ate the bandwidth between two sites.

Fig. 2.

Site Name
Bologna Catania CERN Imperial College Lyon
Storage Element (GBytes)
30
30 10000
80
50
Site Name
Milano NIKHEF NorduGrid Padova RAL Torino
Storage Element (GBytes) 50
70
63
50
50
50
A list of resour es allo ated to the TestBed 1 sites, from whi h the results in
this paper were generated.

Table 1.

4.2 Job Conguration

Initially, all les were pla ed on the CERN Storage Element. Jobs were based on
the CDF use- ase as des ribed in [12. There were six job types, with no overlap

between the set of les ea h job a essed. The total size of the le a essed by
any job type were estimated in [12 and are summarised in Table 2. Ea h set of
les was assumed to be omposed of 10GByte les.
There will be some distribution of jobs ea h site performs. In the simulation,
we modelled this distribution su h that ea h site ran an equal number of jobs
of ea h type ex ept for a preferred job type, whi h ran twi e as often. This job
type was hosen for ea h site based on storage onsiderations; for the repli ation
algorithms to be ee tive, the lo al storage on ea h site had to be able to hold
all the les for the preferred job type.
Data Sample Total Size (GBytes)
Central J=
1200
High pt leptons
200
In lusive ele trons
5000
In lusive muons
1400
High Et photons
5800
Z 0 bb
600

Table 2.

Estimated sizes of CDF se ondary data sets (from [12).

5 Results

A histogram of job duration (left) and the progression of job duration over
ourse of the simulation (right).

Fig. 3.

The left histogram in Figure 3 shows a typi al spread of job duration for a
single job type at a sele ted Computing Element over the ourse of a simulation
run. The large spike near zero is due to the job requesting les that are available
on the lo al site, hen e no time- onsuming le transfers need to take pla e. The
longer durations are due to the job requesting some les not present at the lo al
site. The spread is due to the network load, whi h an vary over time, ae ting
the le transfer times.
The variation of job duration over the simulation is shown in the right histogram in Figure 3 for the same job type and Computing Element as above.
There is learly a large variation in the job duration due to the fa tors already
mentioned, but the general trend is for jobs to be exe uted more qui kly over
time, indi ating the movement toward a more optimal repli a onguration.

Integrated running times for 10000 jobs using ea h a ess pattern and repli a
optimisation algorithm.

Fig. 4.

Further tests were ondu ted simulating 10000 jobs using ea h of the four
algorithms:
1.
2.
3.
4.

No repli ation
Un onditional repli ation, oldest le deleted
Un onditional repli ation, least a essed le deleted
E onomi Model

For ea h repli ation algorithm, ea h of the following four le a ess patterns

(as dened in Se tion 2.2) was tested.
1. Sequential
2. Random

3. Unitary random walk

4. Gaussian random walk
Figure 4 shows the total time to omplete 10000 jobs for ea h of the four
a ess patterns using the four optimisation algorithms.
With no optimisation, the jobs take mu h longer than even the simplest
optimisation algorithm as all the les for every job have to be transferred from
CERN every time a job is run.
The three algorithms where repli ation is ondu ted all show a marked redu tion in the time to exe ute 10000 jobs. This is not surprising as with no
repli ation, all le requests from all jobs must ome from CERN.
The three optimisation algorithms that implement repli ation show similar
performan e for Random, Unitary random walk and Gaussian random walk.
For sequential a ess patterns, the running time is at least 10% faster using
the E onomi Model optimiser than the other optimisers. These results were
expe ted as the E onomi Model assumes a sequential a ess pattern. However,
this an be adjusted to mat h the observed distribution, if needed.

6 Related Work
Re ently there has been great interest in modelling Data Grid environments. A
simulator for modelling omplex data a ess patterns of on urrent users in a
distributed system is found in [13. These studies were mainly ondu ted within
the setting of s ienti experiments su h as the LHC, whi h nally resulted in
the reation of the EU DataGrid proje t [3.
Mi roGrid [18 is a simulation tool for designing and evaluating Grid middleware, appli ations and network servi es for the omputational Grid. Currently,
this simulator does not take data management issues into onsideration. Further
Grid simulators are presented in [11, 6
In [15 an approa h is proposed for automati ally reating repli as in a typi al de entralised Peer-to-Peer network. The goal is to reate a ertain number
of repli as on a given site in order to guarantee some minimal availability requirements.
In Nimrod-G [8, 5 an e onomi model for job s heduling is introdu ed in
where \Grid redits" are assigned to users that are proportional to their level of
priority. In this model, optimisation is a hieved at the s heduling stage of a job.
However, our approa h diers by in luding both optimal repli a sele tion and
automated repli a reation in addition to s heduling-stage optimisation.
Various repli ation and a hing strategies within a simulated Grid environment are dis ussed in [16 and their ombination with s heduling algorithms is
studied in [17. The repli ation algorithms proposed are based on the assumption
that popular les in one site are also popular in other sites. Repli ation from
one site to another is triggered when the popularity of a le over omes a threshold and the destination site is hosen either randomly or by sele ting the least
loaded site. We take a omplementary approa h. Our repli ation algorithms are

used by Grid sites when they need data lo ally and are based on the assumption
that in omputational Grids there are areas (so alled \data hot-spots") where
parti ular sets of data are highly requested. Our algorithms have been designed
to move data les toward \data hot-spots".

7 Con lusions and Future Work

In this paper we des ribed the design of the Grid simulator OptorSim. In parti ular, OptorSim allows the analysis of various repli ation algorithms. The goal
is to evaluate the impa t of the hoi e of an algorithm on the throughput of
typi al Grid jobs. We have hosen two traditional a he management algorithms
(oldest le deletion and least a essed le deletion) and ompared them to a
novel algorithm based on an e onomi model.
We based our analysis on several Grid s enarios with various work loads.
Results obtained from OptorSim suggest that the e onomi model performs at
least as well as traditional methods. In addition, there are spe i realisti ases
where the e onomi model shows marked performan e improvements.
Our future work will extending the simulator by in luding the au tion proto ol proposed in [10. This is motivated by the additional fun tionality of automati repli ation to third party sites, allowing le migration to a urately mat h
demand.

A knowledgements
The authors thank Erwin Laure, Heinz Sto kinger and Ekow Otoo for valuable
dis ussions during the preparation of this paper.
William Bell and David Cameron thank PPARC for funding as part of the
GridPP(EDG) proje t and as an e-S ien e student respe tively; Paul Millar
thanks SHEFC for funding under the S otGRID proje t.

Referen es
1. The ondor proje t. http://www. s.wis .edu/ ondor/.
2. OptorSim - A Repli a Optimiser Simulation. http://grid-data-management.
web. ern. h/grid-data-management/optimisati%on/optor/.
3. The DataGrid Proje t. http://www.eu-datagrid.org.
4. The Globus Proje t. http://www.globus.org.
5. D. Abramson, R. Buuya, and J. Giddy. A Computational E onomy for Grid Computing and its Implementation in the Nimrod-G Resour e Broker. In Future Generation Computer Systems, to appear.
6. K. Aida, A. Takefusa, H. Nakaka, S. Matsuoka, S. Sekigu hi, and U. Nagashima.
Performan e Evaluation Model for S heduling in a Global Computing System.
International Journal of High Performan e Appli ations, 14(3), 2000.
7. W. H. Bell, D. G. Cameron, L. Capozza, P. Millar, K. Sto kinger, and F. Zini. Design of a Query Optimisation Servi e. Te hni al report, CERN, 2002. WP2 - Data
Management, EU DataGrid Proje t. http://edms. ern. h/do ument/337977.

8. R. Buyya, H. Sto kinger, J. Giddy, and D. Abramson. E onomi Models for Management of Resour es in Peer-to-Peer and Grid Computing. In Commer ial Appli ations for High-Performan e Computing, SPIE's International Symposium on
the Convergen e of Information Te hnologies and Communi ations (ITCom 2001),
Denver, Colorado, USA, August 2001.
9. L. Capozza, K. Sto kinger, and F. Zini. Preliminary Evaluation of Revenue Predi tion Fun tions for E onomi ally-Ee tive File Repli ation, June 2002.
10. M. Carman, F. Zini, L. Serani, and K. Sto kinger. Towards an E onomy-Based
Optimisation of File A ess and Repli ation on a Data Grid. In International
Workshop on Agent based Cluster and Grid Computing at International Symposium
on Cluster Computing and the Grid (CCGrid 2002), Berlin, Germany, May 2002.
IEEE Computer So iety Press. Also appears as IRST Te hni al Report 0112-04,
Istituto Trentino di Cultura, De ember 2001.
11. H. Casanova, G. Obertelli an F. Berman, and R. Wolski. The AppLeS Parameter
Sweep Template: User-Level Middleware for the Grid. In Pro . of Super Computing
2002, Dallas, Texas, USA, November 2002.
12. B. T. Human et al. The CDF/D0 UK GridPP Proje t. CDF Internal Note. 5858.
13. I. C. Legrand. Multi-Threaded, Dis rete Event Simulation of Distributed Computing Systems. In Pro . of Computing in High Energy Physi s (CHEP 2000),
Padova, Italy, February 2000.
14. EU DataGrid Proje t. The DataGrid Ar hite ture, 2001.
15. K. Ranganathan, A. Iamnit hi, and I. Foster. Improving Data Availability through
Dynami Model-Driven Repli ation in Large Peer-to-Peer Communities. In Global
and Peer-to-Peer Computing on Large S ale Distributed Systems Workshop, Berlin,
Germany, May 2002.
16. K. Ranganathana and I. Foster. Identifying Dynami Repli ation Strategies for
a High Performan e Data Grid. In Pro . of the International Grid Computing
Workshop, Denver, Colorado, USA, November 2001.
17. K. Ranganathana and I. Foster. De oupling Computation and Data S heduling
in Distributed Data-Intensive Appli ations. In International Symposium of High
Performan e Distributed Computing, Edinburgh, S otland, July 2002. To appear.
18. H. J. Song, X. Liu, D. Jakobsen, R. Bhagwan, X. Zhang, K. Taura, and A. Chien.
The Mi roGrid: a S ienti Tool for Modeling Computational Grids. S ienti
Programming, 8(3):127{141, 2000.