Parallel Biological Sequence Comparison in Large GPU Clusters
summary
CUDAlign is a fine-grained multi-GPU method that compares huge DNA sequences across hundreds of GPUs using parallelogram-shaped parallelism with overlapping compute/communication, an innovative speculation and pruning scheme, and per-block load balancing. It reports state-of-the-art runtime for optimal Smith–Waterman alignments on GPU clusters, includes measured energy consumption, and presents preliminary pruning-on-multiple-GPU results and timings on an IBM Power9 + NVIDIA Volta platform.
AI-written summary · about
chapters (8)
- 0:00Introduction to talk
- 2:31Biological sequence comparison challenges
- 6:50Smith-Waterman algorithm optimization
- 10:50Parallelization strategies for Smith-Waterman
- 15:55Multi-GPU implementation details
- 20:17Performance results and benchmarks
- 24:42Future work and acknowledgments
- 26:32Q&A and closing remarks
related talks
01h 20mFundamental Algorithms in Deep Sequencing II: Mapping/Alignment
18mCompiling Strassen-like Matrix Multiplication Algorithms to Fast CUDA Kernels
26mBuilding a scalable genomic matching algorithm
15mUniversal Checkpointing: A Flexible and Efficient Distributed Checkpointing...
52mC++ Standard Parallelism
22mBeyond Single-Cluster Limits: Scaling GPU Workloads Across Kubernetes With…
42mOpen Source is Better Together- GPU Python Libraries Unite
11mIntroduction to Smith-Waterman – An accurate DNA sequencing algorithm