GPU computing and the tree of life Michael P . Cummings Center for - PowerPoint PPT Presentation

GPU computing and the tree of life Michael P . Cummings Center for Bioinformatics and Computational Biology University of Maryland Institute of Advanced Computer Studies GPU summit 27 October 2014

some domain science context

the great apes

great apes: phylogenetic relationships?

great apes: phylogenetic relationships?

great apes: phylogenetic relationships?

phylogenetic relationships of great apes when subjected to phylogenetic analysis overwhelming evidence supports chimps and humans being each others most closest relatives

number of possible topologies tips unrooted trees 3 1 4 3 5 15 6 105 7 945 8 10,395 9 135,135 10 2,027,025 11 34,459,425 12 654,729,075 13 13,749,310,575 14 316,234,143,225 15 7,905,853,580,625 20 213,643,476,699,771,875

phylogenetic analysis the most accurate methods are . . . . . . . . . . . . T G A G C T T T model-based and involve likelihood A A A C A A A A calculations - maximum likelihood estimation - Bayesian analysis Prob( H | D ) = Prob( D | H ) Prob( H ) _____________ .__Prob( D ) -1574.63624 we can only directly calculate (log likelihood) Prob( D | H )

likelihood calculation peeling algorithm (Felsenstein 1981) does post-order traversal with calculation of partial likelihoods at each node that depend only on its immediate children X ! X ! L ( i ) Prob( x 1 | x 0 , t 1 ) L ( i ) Prob( x 2 | x 0 , t 2 ) L ( i ) 0 ( x 0 ) = 1 ( x 1 ) 2 ( x 2 ) x 1 x 2 x 1 x 2 nonetheless, likelihood calculations are very computationally intensive - t 1 t 2 x 0 O(taxa x sites x rates x states ² ) x

likelihood calculations: majority of computation likelihood related calculations 94.69% nucleotide amino acid 95.72% codon 81.24% GARLI profiling; 11 taxa; 2178 characters

BEAGLE : broad-platform evolutionary analysis general likelihood evaluator an application programming interface ( API ) and high- performance computing library for statistical phylogenetics emphasis is evaluating phylogenetic likelihoods of biomolecular sequence evolution aim is to provide high performance evaluation 'services' to a wide range of phylogenetic software, both Bayesian samplers and maximum likelihood optimizers allows phylogenetic software using the library to make use of optimized hardware such as GPU s

BEAGLE library design goals open-source ( LGPL ) multi-platform support (i.e., Linux, OS X , Windows) low level C API does not explicitly have concept of tree minimize transfer of data support multiple implementations (e.g., CPU , SSE , CUDA , OpenCL) uses dynamic plug-in system support both single and double precision

GPU implementation CPU -side code only used to manage GPU memory allocations and transfers, kernel launches allows client to use CPU in parallel to GPU GPU interface abstraction layer CUDA and OpenCL implementations share same CPU -side code CUDA implementation uses the driver API Parallel Thread Execution ( PTX ) kernels Java Native Interface ( JIT ) compilation templated kernels support arbitrary number of states multiple GPU s supported via client-side partitioning (scales linearly)

gross structure of BEAGLE BEAST JNI wrapper GARLI C API MrBayes implementation manager GPU implementation CPU CUDA interface OpenCL interface CUDA kernels OpenCL kernels BEAGLE

throughput for nucleotide data (4 states) 64 256 16 speedup factor � � � � � � � � � � � � � � � � � � � � � � � � � � � � � 64 GFLOPS � � � � � � 4 � � 16 � � � � � � � 1 � 4 � � � � � 1 100 1,000 1e+04 1e+05 5e+05 3e+06 unique site patterns GPU: AMD Radeon HD 7970 GHz Edition GPU: NVIDIA GeForce GTX 580 (CUDA) GPU: NVIDIA Tesla K20m MIC : Intel Xeon Phi SE10P CPU: Intel Xeon E5−2680 x2 (16 cores) � CPU: Intel Xeon E5−2680 (single core)

throughput for codon data (64 states) 256 1024 64 256 speedup factor GFLOPS � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � 16 64 4 16 1 4 100 1,000 1e+04 6e+04 unique site patterns GPU: AMD Radeon HD 7970 GHz Edition GPU: NVIDIA GeForce GTX 580 (CUDA) GPU: NVIDIA Tesla K20m MIC : Intel Xeon Phi SE10P CPU: Intel Xeon E5−2680 x2 (16 cores) � CPU: Intel Xeon E5−2680 (single core)

MrBayes speedup MrBayes nucleotide model codon model 525 MrBayes SSE 89 64 50 40 � 35 20 23 18 15 16 16 � 17 � GPU: AMD 7970 13 15 MIC : Xeon Phi � 10 CPU: 16 cores � CPU: SSE CPU: standard 4 3.4 3.1 2.6 2.3 1.9 1.6 1.3 1.3 1 double single double single precision

BEAST speedup nucleotide model codon model 889 806 115 64 55 47 44 28 27 25 � 18 16 15 � 16 � GPU: AMD 7970 12 MIC : Xeon Phi 8.6 CPU: 16 cores � � CPU: SSE CPU: standard 4 1.5 1.5 1.4 1.2 1.2 1.0 1 double single double single precision

more than academic academic : having no practical or useful significance Webster’s New Collegiate Dictionary

two recent studies using BEAGLE library

phylogenetics in use: early spread of HIV-1 Faria et al. 2014 The early spread and epidemic ignition of HIV-1 in human populations. Science 346:56-61

phylogenetics in use: 2014 Ebola outbreak Gire et al. 2014 Genomic surveillance elucidates Ebola virus origin and transmission during the 2014 outbreak. Science 345:1369-1372

acknowledgements Daniel Ayres, University of Maryland Peter Beerli, Florida State University Aaron Darling, University of Technology Sydney Mark Holder, University of Kansas John Huelsenbeck, University of California, Berkeley Paul Lewis, University of Connecticut Andrew Rambaut, University of Edinburgh Fredrik Ronquist, Swedish National Museum of Natural History Marc Suchard, University of California, Los Angeles David Swofford, Duke University Derrick Zwickl, University of Arizona Dan Stanzione, Texas Advanced Computing Center Yariv Aridor and Arik Narkis, Intel Israel Altera University Donation Program National Science Foundation