Presentation
Accelerating Sequence Alignment Computations Using Matrix Transformations in Julia
Presenter
DescriptionThe Hamming distance is a fundamental measure of dissimilarity, widely used in error detection, machine learning, and genomic sequence alignment for identifying mismatches in nucleotide or protein sequences. This work introduces two matrix-based implementations—synchronous and asynchronous—for computing Hamming distances in sequence alignment, addressing scalability limitations in traditional vector-based methods. While vector-based approaches are simple and widely used, they lack efficiency for large-scale data. Our asynchronous implementation leverages Julia for sequential task flow and PaRSEC for parameterized task graph execution. CPU computations utilize INT8 GEMM from oneMKL, while GPU implementations employ Tensor/Matrix Core INT8 GEMM from cuBLAS/hipBLAS and 1-bit TensorOps GEMM capabilities from CUTLASS. For bitmask matrix construction on GPUs, we develop a naive CUDA implementation using global memory and an optimized version leveraging shared memory at the warp level, achieving a 5X speedup. The asynchronous matrix-based implementation achieves up to 284X speedup over the vector-based approach on CPUs. On A100 GPUs, the asynchronous GPU-enabled implementation delivers a 15X speedup over the CPU matrix-based approach and improves performance by three orders of magnitude compared to the CPU vector-based method. These results highlight the efficiency and scalability of matrix-based approaches for Hamming distance computation in large-scale applications.
TimeTuesday, June 1716:00 - 16:30 CEST
LocationRoom 5.0B15 & 16
Session Chair
Event Type
Minisymposium
Chemistry and Materials
Engineering
Life Sciences
Physics
Computational Methods and Applied Mathematics