A Methodology for Automatic GPU Kernel Optimization - NECSTTechTalk 4/06/2020
0 likes60 views
The document presents a methodology for automatic GPU kernel optimization and demonstrates its application in optimizing two computationally intensive algorithms: x-drop and Smith-Waterman. The methodology integrates a semi-automatic tool that enhances GPU kernel performance, yielding significant speed-ups, such as over 6.6x compared to seqan for x-drop and over 34x compared to bowtie2 for Smith-Waterman. Overall, the proposed method facilitates effective optimization of GPU performance for high-demand sequence analysis tasks.
![Algorithms Background
14
● Pairwise alignment is one of the most commonly
used workhorses of sequence analysis
● Sequence Alignment is one of the most
computationally expensive steps in genome analysis
86%
SSPACE[1]
70%
PairHMM
GATK[3]
>80%
Bella[2]
[1] Boetzer, Marten, et al. "Scaffolding pre-assembled contigs using SSPACE." Bioinformatics 27.4 (2011): 578-579.
[2] Guidi, Giulia, et al. "BELLA: Berkeley efficient long-read to long-read aligner and overlapper." bioRxiv (2018): 464420.
[3] Sampietro, Davide, et al. "Fpga-based pairhmm forward algorithm for DNA variant calling." 2018 IEEE 29th International Conference on
Application-specific Systems, Architectures and Processors (ASAP). IEEE, 2018.
[4] Li, Heng. "Aligning sequence reads, clone sequences and assembly contigs with BWA-MEM." arXiv preprint arXiv:1303.3997 (2013).
[5] Li, Heng. "Minimap2: pairwise alignment for nucleotide sequences." Bioinformatics 34.18 (2018): 3094-3100.
>85%
BWA-MEM[4]
>85%
minimap2[5]](https://image.slidesharecdn.com/thesispresentation-200604090629/85/A-Methodology-for-Automatic-GPU-Kernel-Optimization-NECSTTechTalk-4-06-2020-14-320.jpg)
![X-Drop Algorithm
19
A
T
T
C
G
G
C
0 -1 -2 -∞
-1 1 -1 -∞
-2 -1 0 -∞
-∞ -∞ -∞
A C G G G G A
● X-Drop [Zhang et al.]
termination offers a great
tradeoff between speed and
accuracy results
● Execution starts from a
common seed to the two
sequences
● Execution stops when the
score drops more than X
from the latest best score](https://image.slidesharecdn.com/thesispresentation-200604090629/85/A-Methodology-for-Automatic-GPU-Kernel-Optimization-NECSTTechTalk-4-06-2020-19-320.jpg)
A Methodology for Automatic GPU Kernel Optimization - NECSTTechTalk 4/06/2020
- 1. POLITECNICO DI MILANO A Methodology for Automatic GPU Kernel Optimization Alberto Zeni
- 2. Context Definition 2 1000x by 2025 40 Years of Microprocessor Trend Data ____________________________________________________ 1980 1990 2000 2010 2020 107 106 105 104 103 102
- 4. Contributions 4 ● We propose a methodology that guides the user to develop highly optimized GPU kernels ● We demonstrate the usefulness of our methodology by implementing it into a semi automatic tool for kernel optimization ● We show the results of the application of our methodology on two highly computationally intensive algorithms
- 5. Methodology application 5 Unoptimized source code Roofline Generator Roofline and Performance Analyzer Optimized source code Optimization Flow Compiler Optimizer Source Code Parser
- 6. Methodology application 6 Unoptimized source code Roofline Generator Roofline and Performance Analyzer Optimized source code Optimization Flow Compiler Optimizer Source Code Parser
- 8. Roofline Model Adaptation 8 ● Model built on the characteristics of the GPU and algorithm executed and independent to the algorithm implementation = number of iterations = gpu cores frequency = number of operations to be computed at iteration i = number of blocks = number of scheduled threads per block = number of integer cores = = number of streaming multiprocessors = maximum number of blocks per streaming multiprocessor =
- 9. Methodology application 9 Unoptimized source code Roofline Generator Roofline and Performance Analyzer Optimized source code Optimization Flow Compiler Optimizer Source Code Parser
- 10. Methodology application 10 Unoptimized source code Roofline Generator Roofline and Performance Analyzer Optimized source code Optimization Flow Compiler Optimizer Source Code Parser
- 11. Source Code Parser 11 ● Automatically unrolls loops if possible ● Automatically changes the memory hierarchy ● Automatically changes the number of scheduled threads ● Automatically changes the number of scheduled blocks ● Creates a report of the optimizations that can be applied manually
- 12. Methodology application 12 Unoptimized source code Roofline Generator Roofline and Performance Analyzer Optimized source code Optimization Flow Compiler Optimizer Source Code Parser
- 13. Methodology application 13 Unoptimized source code Roofline Generator Roofline and Performance Analyzer Optimized source code Optimization Flow Compiler Optimizer Source Code Parser
- 14. Algorithms Background 14 ● Pairwise alignment is one of the most commonly used workhorses of sequence analysis ● Sequence Alignment is one of the most computationally expensive steps in genome analysis 86% SSPACE[1] 70% PairHMM GATK[3] >80% Bella[2] [1] Boetzer, Marten, et al. "Scaffolding pre-assembled contigs using SSPACE." Bioinformatics 27.4 (2011): 578-579. [2] Guidi, Giulia, et al. "BELLA: Berkeley efficient long-read to long-read aligner and overlapper." bioRxiv (2018): 464420. [3] Sampietro, Davide, et al. "Fpga-based pairhmm forward algorithm for DNA variant calling." 2018 IEEE 29th International Conference on Application-specific Systems, Architectures and Processors (ASAP). IEEE, 2018. [4] Li, Heng. "Aligning sequence reads, clone sequences and assembly contigs with BWA-MEM." arXiv preprint arXiv:1303.3997 (2013). [5] Li, Heng. "Minimap2: pairwise alignment for nucleotide sequences." Bioinformatics 34.18 (2018): 3094-3100. >85% BWA-MEM[4] >85% minimap2[5]
- 15. 15 2 2 0 0 0 0 4 0 A A T G A T T C ● Optimal algorithm for local sequence alignment ● Perform the alignment by computing a matrix called Alignment Matrix ● The algorithm has a fixed score it two characters do or do not match ● The score of a cell is determined by following dependencies on the previously computed cells Smith-Waterman Algorithm Match = 2 Mismatch = -2
- 16. Smith-Waterman Algorithm 16 ● Optimal algorithm for local sequence alignment ● Execution times scale up to the length of the aligned sequences A G G G T C A A 0 0 0 0 0 0 0 0 0 0 1 0 0 0 1 0 0 1 0 0 0 0 0 0 2 1 0 0 0 0 0 0 0 1 3 1 0 0 0 0 0 0 1 2 2 0 0 0 0 1 0 0 0 1 0 0 1 1 0 0 0 0 0 0 1 0 0 0 1 0 0 1 0 1 0 0 0 1 0 0 1 A C C T A G G A
- 17. 17 1 -1 -3 -4 -1 0 -2 -3 A C G G A T T C ● Optimal algorithm for global sequence alignment ● Perform the alignment by computing a matrix called Alignment Matrix ● The algorithm has a fixed score it two characters do or do not match ● The score of a cell is determined by following dependencies on the previously computed cells The Needleman-Wunsch Algorithm Match = 2 Mismatch = -2
- 18. 18 The Needleman-Wunsch Algorithm ● Optimal algorithm for global sequence alignment ● Execution time correlated to the length of the aligned sequences ● Inefficient if the two sequences do not align A T T C G G C 0 -1 -2 -3 -4 -5 -6 -7 -1 1 -1 -3 -4 -5 -6 -5 -2 -1 0 -2 -4 -5 -6 -7 -3 -3 -2 -1 -3 -5 -6 -7 -4 -4 -2 -3 -2 -4 -6 -7 -5 -5 -4 -1 -2 -1 -3 -7 -6 -6 -6 -3 0 -1 0 -4 -7 -7 -5 -5 -2 -1 -2 -1 A C G G G G A
- 19. X-Drop Algorithm 19 A T T C G G C 0 -1 -2 -∞ -1 1 -1 -∞ -2 -1 0 -∞ -∞ -∞ -∞ A C G G G G A ● X-Drop [Zhang et al.] termination offers a great tradeoff between speed and accuracy results ● Execution starts from a common seed to the two sequences ● Execution stops when the score drops more than X from the latest best score
- 20. X-Drop Algorithm 20 ● X-Drop computation of each cell is the same of NW A T T C G G C 0 -1 -1 A C G G G G A X=2 MAX = 0 Alignment = +1 All penalties = -1
- 21. X-Drop Algorithm 21 ● X-Drop computation of each cell is the same of NW A T T C G G C 0 -1 -2 -1 1 -2 A C G G G G A X=2 MAX = 1 Alignment = +1 All penalties = -1
- 22. X-Drop Algorithm 22 ● If the score of the cell is below more than X from MAX then the cell is flagged with -infA T T C G G C 0 -1 -2 -1 1 -2 A C G G G G A X=2 MAX = 0 Alignment +1 All penalties -1
- 23. X-Drop Algorithm 23 ● Once an antidiagonal has been computed the global maximum is updatedA T T C G G C 0 -1 -2 -1 1 -2 A C G G G G A X=2 MAX = 1 Alignment = +1 All penalties = -1
- 24. X-Drop Algorithm 24 ● X-Drop computation of each cell is the same of NW A T T C G G C 0 -1 -2 -3 -1 1 -1 -2 -1 -3 A C G G G G A X=2 MAX = 1 Alignment = +1 All penalties = -1
- 25. X-Drop Algorithm 25 ● -3 is below MAX - X so the cell is flagged with -inf ● MAX remained the sameA T T C G G C 0 -1 -2 -∞ -1 1 -1 -2 -1 -∞ A C G G G G A X=2 MAX = 1 Alignment = +1 All penalties = -1
- 26. X-Drop Algorithm 26 ● Again we compute the cells normally ● Now -2 is below MAX - X so the cell is flagged with -inf ● MAX remained the same A T T C G G C 0 -1 -2 -∞ -1 1 -1 -2 -2 -1 0 -∞ -2 A C G G G G A X=2 MAX = 1 Alignment = +1 All penalties = -1
- 27. X-Drop Algorithm 27 ● Again we compute the cells normally ● Now -2 is below MAX - X so the cell is flagged with -inf ● MAX remained the same A T T C G G C 0 -1 -2 -∞ -1 1 -1 -∞ -2 -1 0 -∞ -∞ A C G G G G A X=2 MAX = 1 Alignment = +1 All penalties = -1
- 28. X-Drop Algorithm 28 ● Again we compute the cells normally ● Now all the cells scores are below MAX - X A T T C G G C 0 -1 -2 -∞ -1 1 -1 -∞ -2 -1 0 -2 -∞ -∞ -2 A C G G G G A X=2 MAX = 1 Alignment = +1 All penalties = -1
- 29. X-Drop Algorithm 29 ● Again we compute the cells normally ● Now all the cells scores are below MAX - X ● The execution ends A T T C G G C 0 -1 -2 -∞ -1 1 -1 -∞ -2 -1 0 -∞ -∞ -∞ -∞ A C G G G G A X=2 MAX = 1 Alignment = +1 All penalties = -1
- 30. X-Drop search space 30 X-Drop Banded Comparison between the search space of different algorithms. ● X-Drop is a very efficient alignment heuristic to align genomic sequences, especially if the two sequences do not align ● It offers a significant improvement when compared to Banded and classic NW, as computation can be cut earlier if needed. NW
- 31. GPU implemented optimizations 31 ● The two algorithms follow the same computational pattern ● We started with a simple implementation of the algorithms using a single thread and a single block ● We followed our methodology with the help of our tool to optimize the algorithms at different levels and introduce Inter and Intra Parallelism 1 -1 -3 -4 -1 0 -2 -3 A C G G A T T C A C G G A T T C
- 32. Intra Level Parallelism 32 ● Parallel computation of the anti-diagonals ● Each GPU thread is assigned to compute a single cell as our methodology suggested ● Anti-diagonals split in different segments to align sequences of any length
- 33. Inter Level parallelism 33 ● Parallel execution of the alignments with multiple blocks ● Each block has an alignment assigned
- 34. GPU memory optimizations 34 ● To ensure coalesced memory access one of the sequences is stored backwards on the GPU
- 35. 35 Evaluation Settings Benchmarked Applications: ● SeqAn highly optimized version of X-Drop ● ksw2: CPU SIMD Z-drop ● Bowtie2: Smith-Waterman ● CUDASW++ 3.0: Smith-Waterman GPU + CPU SIMD
- 36. Evaluation Settings 36 Platforms: ● Intel Haswell Nodes ● Intel Skylake Nodes ● IBM Power 9 Nodes ● GPU
- 42. X-drop GPU and SeqAn Comparison 42 2x 6x
- 43. X-drop GPU and ksw2 Comparison 43 1.5x 120x
- 44. Conclusions 44 A methodology for automatic GPU Kernel Optimization and its implementation inside a tool for automatic kernel optimization We applied our methodology to two highly computational intensive algorithms Optimized GPU X-drop Implementation with: ● More than 6.6x speed-up with respect to SeqAn ● More than 120x speed-up with respect to ksw2 Optimized GPU Smith-Waterman Implementation with: ● More than 34x speed-up with respect to Bowtie2 ● More than 3x speed-up with respect to CUDASW++ 3.0
- 45. Thank you for your attention A Methodology for Automatic GPU Kernel Optimization Alberto Zeni