and Argonne National Laboratory Computation Institute...

1

Parallel scripting with Swift for applications at the petascale and beyond

VecPar PEEPS Workshop

Berkeley, CA – June 22, 2010

Michael Wilde – wilde@mcs.anl.govComputation Institute, University of Chicago

and Argonne National Laboratory

www.ci.uchicago.edu/swift

• Many applications need loosely coupled scripting

• Swift harnesses parallel & distributed resources through a simple scripting language

• Productivity gains by enabling use of more powerful systems with less concern for the mechanics

2

Problems addressed by Swift

Modeling uncertainty for CIM‐EARTH

Parallel AMPL workflow by Joshua Elliott, Meredith Franklin, Todd Munson, Allan Espinosa.

4

Fast Ocean Atmosphere Model (MPI)

NCAR

Manual config, execution,

bookkeeping

VDS on Teragrid

Automated

Visualization courtesy Pat Behling and Yun Liu, UW

MadisonWork ofVeronica Nefedovaand Rob Jacob, Argonne

Problem: Drug screening at APS

5

2M+ ligands

(Mike Kubal, Benoit Roux, and others)

(B)

O(Millions)of drug

candidates

O(tens)of fruitful

candidates for wetlab &

APS

start

report

DOCK6Receptor

(1 per protein:defines pocket

to bind to)

ZINC3-D

structures

ligands complexes

NAB scriptparameters

(defines flexibleresidues, #MDsteps)

Amber Score:1. AmberizeLigand3. AmberizeComplex5. RunNABScript

end

BuildNABScript

NABScript

NABScript

Template

Amber prep:2. AmberizeReceptor4. perl: gen nabscript

FREDReceptor

(1 per protein:defines pocket

to bind to)

Manually prepDOCK6 rec file

Manually prepFRED rec file

1 protein(1MB)

6 GB2M

structures(6 GB)

DOCK6FRED ~4M x 60s x 1 cpu~60K cpu-hrs

Amber ~10K x 20m x 1 cpu~3K cpu-hrs

Select best ~500

~500 x 10hr x 100 cpu~500K cpu-hrsGCMC

PDBprotein

descriptions

Select best ~5KSelect best ~5K

6Work of Andrew Binkoaski and Michael Kubal

Problem: preprocessing and analysisof neuroscience experiments

ManyData Files:

3a.h

align_warp/1

3a.i

3a.s.h

softmean/9

3a.s.i

3a.w

reslice/2

4a.h

align_warp/3

4a.i

4a.s.h 4a.s.i

4a.w

reslice/4

5a.h

align_warp/5

5a.i

5a.s.h 5a.s.i

5a.w

reslice/6

6a.h

align_warp/7

6a.i

6a.s.h 6a.s.i

6a.w

reslice/8

ref.h ref.i

atlas.h atlas.i

slicer/10 slicer/12 slicer/14

atlas_x.jpg

atlas_x.ppm

convert/11

atlas_y.jpg

atlas_y.ppm

convert/13

atlas_z.jpg

atlas_z.ppm

convert/15

ManyApplication Programs:

Automated image registration for spatial normalization

reorientRun

reorientRun

reslice_warpRun

random_select

alignlinearRun

resliceRun

softmean

alignlinear

combinewarp

strictmean

gsmoothRun

binarize

reorient/01

reorient/02

res li c e_w arp/22

alignl inear/03 al ignl inear/07alignl inea r/11

reorien t/05

reorient/06

res lic e_w ar p/23

reo rien t/09

reorient/10

res l ic e_w arp/24

reorien t/25

reorient/51

res l ic e_warp/26

reorient/27

reo rient/52

r es l ic e_warp/28

reorient/29

reo rient/53

res l ic e_warp/30

reorien t/31

reorient/54

res lic e_w arp/32

reorient/33

r eorient/55

resl i c e_warp/34

reorient/35

reor ient/56

resl i ce_warp/36

reor ient/37

reorien t/57

resl ice_warp /38

resl i ce /04 resl ice/08res l i ce/12

gsm ooth/41

s tr ic tmean/39

gs mooth/42gs mooth /43gs mooth/44 gsm oo th/45 gs mooth/46 gs m ooth /47 gs mooth/48 gs m ooth /49 gsmooth/50

softm ean/13

a lign l inear/17

com binewarp/21

binari ze/40

reorient

reorient

alignlinear

reslice

softmean

alignlinear

combine_warp

reslice_warp

strictmean

binarize

gsmooth

AIRSN workflow expanded:AIRSN workflow:

Swift programs

• A Swift script is a set of functions– Atomic functions wrap & invoke application programs (on parallel

compute nodes)

– Composite functions invoke other functions (run in Swift engine)

• Data is typed as composable arrays and structures of files and simple scalar types (int, float, string)

• Collections of persistent file structures are mapped into this data model as arrays and structures

• Variables are single assignment

• Expressions and statements are executed in data‐flowdependency order and concurrency

• Members of datasets can be processed in parallel

• Provenance is gathered as scripts execute9

A simple Swift scriptTo run the Image Magick app “convert”:1 convert ‐rotate 180 $in $out

2 type imagefile { } // Declare a “file” type.3

4 app (imagefile output) rotate (imagefile input) {5 {6 convert "‐rotate" 180 @input @output ;7 }8

9 imagefile image <"m101.2010.0601.jpg">;10 imagefile newimage <"output.jpg">;11

12 newimage = rotate(image);10

Execution is driven by data flow

1 (int result) myproc (int input)2 {3 j = f(input); 4 k = g(input);5 result = j + k;6 }

7 j=f() and k=g() are computed in parallel.

8 This parallelism is automatic, based on futures;9 Works recursively down the scripts’s call graph.

11

Parallelism via foreach { }1 type imagefile; // Declare a “file” type.2

3 app (imagefile output) rotate (imagefile input) {4 convert "‐rotate" "180" @input @output;5 }6

7 imagefile observations[ ] <simple_mapper; prefix=“m101‐raw”>;8 imagefile flipped[ ] <simple_mapper; prefix=“m101‐flipped”>;9

10

11

12 foreach obs,i in observations {13 flipped[i] = rotate(obs); 14 }

12

Name outputs based on index

Process all dataset members in parallel

Map inputs from local directory

Many domains process structured datasets

ManyData Files:

3a.h

align_warp/1

3a.i

3a.s.h

softmean/9

3a.s.i

3a.w

reslice/2

4a.h

align_warp/3

4a.i

4a.s.h 4a.s.i

4a.w

reslice/4

5a.h

align_warp/5

5a.i

5a.s.h 5a.s.i

5a.w

reslice/6

6a.h

align_warp/7

6a.i

6a.s.h 6a.s.i

6a.w

reslice/8

ref.h ref.i

atlas.h atlas.i

slicer/10 slicer/12 slicer/14

atlas_x.jpg

atlas_x.ppm

convert/11

atlas_y.jpg

atlas_y.ppm

convert/13

atlas_z.jpg

atlas_z.ppm

convert/15

ManyApplication Programs:

Swift Data Mappingtype Study {

Group g[ ]; }

type Group { Subject s[ ];

}

type Subject { Volume anat; Run run[ ];

}

type Run { Volume v[ ];

}

type Volume { Image img; Header hdr;

}

On-DiskDataLayout Swift’s

in-memory data model

Mapping functionor script

Mapping functionor script

Swift app function“predict()”

Swift app function“predict()”

tt

seqseq dt

dtloglog

PSim applicationPSim application

pgpg

To run:psim –s 1ubq.fas –pdb p \

–temp 100.0 –inc 25.0 >log

In Swift code:

app (PDB pg, File log) predict (Protein seq,Float temp, Float dt)

{psim "-s" @pseq.fasta "-pdb" @pg

"–temp" temp ”-inc" dt;}

Protein p <ext; exec="Pmap", id="1ubq">;ProtGeo structure;TextFile log;

(structure, log) = predict(p, 100., 25.);

Fastafile

Fastafile

Application: Protein structure prediction

Encapsulation is the key to transparent distribution, parallelization, and provenance

foreach sim in [1:1000] {(structure[sim], log[sim]) = predict(p, 100., 25.);

}result = analyze(structure)

1000predict()

calls

Analyze()

Parallelism via foreach { }

Application: 3D Protein structure prediction

1. type Fasta; // Primary protein sequence file in FASTA format2. type SecSeq; // Secodary structure file3. type RamaMap; // “Ramachandra” mapping info files4. type RamaIndex;5. type ProtGeo; // PDB‐format file – protein geometry: 3D atom coords6. type SimLog;7.8. type Protein { // Input file struct to protein simulator9. Fasta fasta; // sequence to predict structure of10. SecSeq secseq; // Initial secondary structure to use11. ProtGeo native; // 3D structure from experimental data when known 12. RamaMap map;13. RamaIndex index;14. }15.16. type PSimCf { // Science configuration parameters to simulator17. float st;18. float tui;19. float coeff;20. }21.22. type ProtSim { // Output file struct from protein simulator23. ProtGeo pgeo;24. SimLog log;25. } 17

Protein structure prediction

1. app (ProtGeo pgeo) predict (Protein pseq)2. {3. PSim @pseq.fasta @pgeo;4. }5.6. (ProtGeo pg[ ]) doRound (Protein p, int n) {7. foreach sim in [0:n‐1] {8. pg[sim] = predict(p);9. }10. }11.12. Protein p <ext; exec="Pmap", id="1af7">;13. ProtGeo structure[ ];14. int nsim = 10000;15. structure = doRound(p, nsim);

18

Protein structure prediction

1 (ProtSim psim[ ]) doRoundCf (Protein p, int n, PSimCf cf) {2 foreach sim in [0:n‐1] {3 psim[sim] = predictCf(p, cf.st, cf.tui, cf.coeff );4 }5 }

6 (boolean converged) analyze( ProtSim prediction[ ], int r, int numRounds)7 {8 if( r == (numRounds‐1) ) {9 converged = true;10 }11 else {12 converged = test_convergence(prediction);13 }14 }

19

Protein structure prediction

1. ItFix( Protein p, int nsim, int maxr, float temp, float dt)2. {3. ProtSim prediction[ ][ ];4. boolean converged[ ];5. PSimCf config;6.7. config.st = temp;8. config.tui = dt;9. config.coeff = 0.1;10.11. iterate r {12. prediction[r] =13. doRoundCf(p, nsim, config);14. converged[r] =15. analyze(prediction[r], r, maxr);16. } until ( converged[r] );17. }

20

Protein structure prediction

1. Sweep( )2. {3. int nSim = 1000;4. int maxRounds = 3;5. Protein pSet[ ] <ext; exec="Protein.map">;6. float startTemp[ ] = [ 100.0, 200.0 ];7. float delT[ ] = [ 1.0, 1.5, 2.0, 5.0, 10.0 ];8. foreach p, pn in pSet {9. foreach t in startTemp {10. foreach d in delT {11. ItFix(p, nSim, maxRounds, t, d);12. }13. }14. }15. }16.17. Sweep();

21

10 proteins x 1000 simulations x3 rounds x 2 temps x 5 deltas

= 300K tasks

Submit host(Laptop, login host,…)

Workflowstatus

and logs

?????

Computenodes

f1

f2

f3

a1

a2

Data server

f1 f2 f3

Provenancelog

scriptAppa1

Appa2

sitelist

applist

Filetransport

Filetransport

CloudsClouds

UsingSwift

Swift is a self‐contained application with cluster and grid client code:Dowload, untar, and run

Small, fast, localmemory-based filesystems

Falkon client(load

balancing) Sharedglobal

filesystem

Swift scriptFalkon services on

BG/P IO Processors

BG/P Processor sets

Architecture for petascale scripting

Collective data management is critical for petascale

• Applies “scatter/gather” concepts at the file management level

• Seeks to avoid contention, maximize parallelism and use petascale interconnects– Broadcast common files to compute nodes– Place per‐task data on local (RAM) FS– Gather output into larger sets (time/space)– Aggregate small local FS’s into large striped FS

• Still in research phase: paradigm and architecures

24

Collective data management

Global FS

IFSNodeIFS

. . .

CNLFS

CNLFS

. . .

CNLFS

CNLFS

. . .

IFSNodeIFS

DistributorCollector

55

44

33

22

11

Performance: Molecular dynamics on BG/P

26935,803 DOCK jobs with Falkon on BG/P in 2 hours

Performance: SEM for fMRI on Constellation

27418K SEM tasks with Swift/Coasters on Ranger in 41 hours

Performance: Proteomics on BG/P

284,127 PTMap jobs with Swift/Falkon on BG/P in 3 minutes

Scaling the many‐task model

29

Compute unit

Compute unit

Client (master) application

Compute unit

graph executor

Mastergraph

executor

Compute unit

graph executor

Compute unit

graph executor

Compute unit

graph executor

Virtual data store

Global persistent storage

Extreme-scale computing complex

Ultra-fast Message queues

Scaling many‐task computing

• ADLB: tasks can be lightweight functions– Retains RPC model of input‐process‐output

– Fast, distributed, asynchronous load balancing

• Multi‐level task manager– Must scale to massive computing complexes

• Transparent distributed management of local storage– Leverage local filesystems (RAM), aggregate,

make access more transparent through DHT methods 30

Conclusion: Motivation for Swift

• Enhance scientific productivity– Location – and paradigm – independence:

Same scripts run on workstations, clusters, clouds, grids, and petascale supercomputers

– Automation of dataflow, resource selection and error recovery

• Enable and motivate collaboration– Community libraries of techniques, protocols,

methods– Designed for recording the provenance of all data

produced to facilitate scientific processes

• Swift is a parallel scripting system for Grids and clusters– for loosely‐coupled applications ‐ application and utility programs linked

by exchanging files

• Swift is easy to write: simple high‐level C‐like functional language– Small Swift scripts can do large‐scale work

• Swift is easy to run: contains all services for running Grid workflow ‐in one Java application– Untar and run – acts as a self‐contained Grid client

• Swift is fast: Karajan provides Swift a powerful, efficient, scalable and flexible execution engine.– Scaling close to 1M tasks – .5M in live science work, and growing

• Swift usage is growing:– applications in neuroscience, proteomics, molecular dynamics,

biochemistry, economics, statistics, and more.

To learn more and try Swift…

33

• www.ci.uchicago.edu/swift– Quick Start Guide:

• http://www.ci.uchicago.edu/swift/guides/quickstartguide.php

– User Guide:• http://www.ci.uchicago.edu/swift/guides/userguide.php

– Introductory Swift Tutorials:• http://www.ci.uchicago.edu/swift/docs/index.php

http://www.ci.uchicago.edu/swift

34IEEE COMPUTER, Nov 2009

Acknowledgments• Swift effort is supported in part by NSF grants OCI‐721939, OCI‐0944332, and PHY‐

636265, NIH DC08638, and the UChicago/Argonne Computation Institute• The Swift team (present and former):

– Ben Clifford, Allan Espinosa, Ian Foster, Mihael Hategan, Ioan Raicu, Sarah Kenny, Mike Wilde, Justin Wozniak, Zhao Zhang, Yong Zhao

• Java CoG Kit used by Swift developed by:– Mihael Hategan, Gregor Von Laszewski, and many collaborators

• Falkon software– developed by Ioan Raicu and Zhao Zhang

• ZeptoOS– Kamil Iskra, Kazutomo Yoshii, and Pete Beckman

• Scientific application collaborators and users– U. Chicago Open Protein Simulator Group (Karl Freed, Tobin Sosnick, Glen Hocky, Joe

Debartolo, Aashish Adhikari)– U.Chicago Radiology and Human Neuroscience Lab, (Dr. S. Small)– SEE/CIM‐EARTH/Econ: Joshua Elliott, Meredith Franklin, Todd Muson, Tib Stef‐Praun– PTMap: Yingming Zhao, Yue Chen

35