varre_biomanycores_bosc2009

Biomanycores, a repository of interoperableopen-source code for many-cores bioinformatics

Jean-Stephane Varre, Stephane Janot, Mathieu [email protected]

Sequoia BioinformaticsLIFL – UMR CNRS 8022 – Universite Lille 1, France

INRIA Lille Nord-Europe, France

June 2009

J.-S. Varre, S. Janot, M. Giraud (LIFL) Biomanycores June 2009 1 / 20

Outline

High-performance computing

Graphical Processing Units and bioinformatics

biomanycores.orgI aim of the projectI what has been done ?I future developments


High Performance Bioinformatics – Manycores

1970 – 2002:Moore’s law =increasing frequencies

problems:power consumption,heat dissipation here

from now on: Moore’s law continues with multiple coresI from multicores: dual-cores, quad-cores, octo-cores...I to manycores:

F Graphic processing units (GPUs)Nvidia GTX 285 ⇒ 30× 8 cores, 1.2 GHz, 40 (×8) GFlops

F convergence CPU-GPU: Intel Larrabee


High Performance Bioinformatics – Manycores

GPGPU = General-Purpose computation on GPU

until 2007: tweaking graphics primitives

2007: Nvidia CUDA

2009: OpenCL (Khronos Group)I dec 08: 1.0 specificationI may 09: beta release of a Nvidia compilerI AMD/ATI compiler coming soon

⇒ portable manycores applications ?

With GPGPU...

10× / 100× peak speed-up, low costs ($50–$500)

even with loss due to parallelism, 10× speed-up is possible

(relatively) easy with CUDA / OpenCL, requires some learning


GPU + BioinformaticsMethods

“Graphical” GPGPU (2005/06):

speed-upRAxML up to 2× Charalambous et al. 2005

ClustalW up to 7× Liu et al. 2006

CUDA (since 2007):

speed-upmummerGPU up to 10× Schatz et al. 2007

Smith-Waterman up to 15× Manavski and Valle 2008Neighbor-Joining up to 26× Liu et al. 2009

RNAfold up to 17× Risk and Lavenier 2009

∼ 10 papers between 2007 and 2009


GPU + BioinformaticsSpecific Bioinformatics HPC Events

HiComb (IEEE Workshop on High Performance Computational Biology)since 2002

in conjunction with IPDPS [may 09, Roma]

PBC (Parallel Bio-Computing Workshop)since 2005, every two years

in conjunction with PPAM [sept 09, Wroclaw]

HiBi (Workshop on High Performance Computational Systems Biology)

[oct 09, Trento]


Sequoia BioinformaticsLIFL, INRIA, Universite Lille 1, France

H. Touzet’s group, 14 people (including 5 PhD students)

Large-scale sequence analysis

Sequence comparisons, seed-based heuristics

RNA, transcription factors, NRPS

High-Performance BioinformaticsI SIMD flexible read mapper (L. Noe, M. Gırdea)I GPU PWM scan / P-value (22× – 77× on a GTX 280)I GPU ADP (6.1× – 22.8× on a GTX 280, with U. Bielefeld)I GPU & bit-parallelism pattern matching (ongoing)I Supported by NVIDIA (Professor Partnership, 2009)


GPU + Position-Weight Matrices (PWM)Parallel Position Weight Matrices Algorithms. M. Giraud and J.-S. Varre. ISPDC’09

PWMs are used for modeling transcriptionfactor binding sites, transcription start sites,protein domains, . . .

score threshold or P-value computation:requires to enumerate words

occurrences: requires to scan quickly a verylong sequence

WebLogo 3.0

0.0

1.0

2.0

bits

CTAACTCATT

G5A

CTGG

GC

ATACT

100x

10x

1x

35 40 45 50 55 60 65 70

Spee

dup

Matrix length

CPU (one thread)GeForce 8800

GTX 280GTX 280 (+ atomic)

25x

20x

15x

10x

5x

0 10 20 30 40 50 60 70 80 90

Spee

dup

Matrix length

CPU (one thread)GeForce 8800

GTX 280


HPC Bioinformatics for human beings ?

Research in High-Performance ComputingI nice ideas, nice papersI but not always exploited

A few HPC bioinformatics frameworks projects...

⇒ far from everyday usage of bioinformaticians and biologists


www.biomanycores.org

1. Share OpenCL code= public repository, open-source

2. Make it easy= Bio∗ integration

3. Benchmarkalgorithms, implementations, hardware



1. Share OpenCL code (currently CUDA)

= public repository, open-source




Already included projects

SWcuda – Smith-Waterman protein alignmentI CRIBI Genomics, University of Padova, ItalyI S. A. Manavski, G. Valle, CUDA compatible GPU cards as efficient hardware

accelerators for Smith-Waterman sequence alignment, BMC Bioinformatics 2008,9(S2):S10

pknotsRG – pseudonots of an RNA sequenceI Universitat Bielefeld, GermanyI J. Reeder, P. Steffen, R. Giegerich, pknotsRG: RNA pseudoknot folding including

near-optimal structures and sliding windows, Nucl. Acids. Res., 2007

cudaPWM – scan a PWM against a DNA sequenceI Sequoia, LIFL, INRIA, Universite Lille 1I M. Giraud, J.-S. Varre, Parallel Position Weight Matrices Algorithms, ISPDC’09

Interfaces to BioJava 1.6, BioPerl 1.52, and Biopython 1.50b


Biopython + CRIBI SW

from Bio i m p o r t SeqIO

from Biomanycores i m p o r t PadovaSW

bank = SeqIO . parse ( open ( ” u n i p r o t−s t a r t . f a ” ) , ” f a s t a ” )

f o r query i n SeqIO . parse ( open ( ” p r o t 6 4 . f a ” ) , ” f a s t a ” ) :handle = PadovaSW . run ( query , bank )result = PadovaSW . SWParser ( ) . parse ( )p r i n t result


Biopython + CRIBI SWTests on a GeForce 8800

biopython$ time python sw-demo.py cuda

** cd ../bin/ ; ./swcuda config.gpu ../tmp/swcuda.fa ../tmp/swcuda.bank

** 1.846s

12098 results...

[(84.0, 0, 0, ’sp|P30350|ADH1_ANAPL’), (81.0, 0, 0, ’sp|P23991|ADH1_CHICK’), (81.0, 0, 0, ’sp|P72324|ADHI_RHOS4’), (79.0, 0, 0, ’sp|P29274|AA2AR_HUMAN’), ...]

real 2.81 user 1.79 sys 0.27

biopython$ time python sw-demo.py cpu

** cd ../bin/ ; ./swcuda config.cpu ../tmp/swcuda.fa ../tmp/swcuda.bank

** 16.604s

12098 results...

[(84.0, 0, 0, ’sp|P30350|ADH1_ANAPL’), (81.0, 0, 0, ’sp|P23991|ADH1_CHICK’), (81.0, 0, 0, ’sp|P72324|ADHI_RHOS4’), (79.0, 0, 0, ’sp|P29274|AA2AR_HUMAN’), ...]

real 17.57 user 16.42 sys 0.14

10× – 15× paper speedup (BMC Bioinformatics 2008, 9S2)

8.7× application speedup

6.2× final speedup (including Biopython/Biomanycores)


BioPerl + CRIBI SW

BioPerl tutorial

use Bio : : Tools : : pSW ;

$factory = new Bio : : Tools : : pSW ( ’−m a t r i x ’=> ’ blosum62 . b l a ’ , ’−gap ’←↩=>12, ’−e x t ’=>2) ;

$factory−>align_and_show ( $seq1 , $seq2 , STDOUT ) ;$aln = $factory−>pairwise_alignment ( $seq1 , $seq2 ) ;

With biomanycores

use Bio : : SeqIO ;use Biomanycores : : PadovaSW ;

$factory = PadovaSW−>new ( ) ;

$factory−>swcuda ( $inputseq , $bank ) ;@r = $factory−>parse_result ( ) ;


BioJava + PWM

i m p o r t org . biojavax . bio . seq . RichSequence ;i m p o r t org . biojava . bio . dp . SimpleWeightMatrix ;. . .i m p o r t org . biomanycores . bio . pwm . ∗ ;. . .{

LillePWMScan scanner = new LillePWMScan ( launcher ) ;

// r e a d t h e s e q u e n c eRichSequenceIterator it = n u l l ;BufferedReader in1 = new BufferedReader ( new FileReader ( args [ 1 ] ) ) ;it = RichSequence . IOTools . readFastaDNA ( in1 , n u l l ) ;RichSequence query = it . nextRichSequence ( ) ;

// r e a d a w e i g h t m a t r i xSimpleWeightMatrix pwm = PFMParser . PARSER . get ( args [ 2 ] , alph , ”ACGT” ) ;

// scan t h e s e q u e n c eList<PWMHit> al = scanner . scan ( query , pwm , threshold ) ;

}


Challenges

Differents APIs, different philosophiesI BioJava : no external program execution ?I Object representation (alignments)I Object existence (PWM)

Minimal modifications to the source code of applicationsI CribiSW : command-line arguments

Real-world pipelines ?I Bio∗ are not HPC frameworksI Succession of several programs

Usage: requires CUDA / OpenCL SDK


Licenses

Projects must have an open-source licence

Bio∗ interfaces : same license than mother APII BioJava: LGPL 2.1I BioPerl: Perl artistic licenseI Biopython: Biopython license



1. Share OpenCL code (currently CUDA)

= public repository, open-source

⇒ bring new projects


⇒ integrate new projects

⇒ improve current interfaces


⇒ think !


go back

varre_biomanycores_bosc2009

Technology

giraud lifl ebiomanycoresjune

giraud lifl e biomanycoresjune

sequoia bioinformatics

stphane janot

bmc bioinformatics

cuda opencl

universit lille

opencl code