Abstract:
Data movement between memory and the CPU is a bottleneck in data-intensive applications. The cause of this problem is the need for the computing unit to access data frequently in memory through a limited-bandwidth and high-latency memory bus. Processing-in-memory (PIM) architectures solve this problem by bringing computing units closer to the memory on the same memory chip. The aim of this thesis is to show that genome sequence analysis can be effectively accelerated using PIM architectures. We perform high-throughput read pair alignment using the Needleman-Wunsch, Smith-Waterman-Gotoh, GenASM, and the Wave Front Alignment algorithms on a real PIM architecture. The performance was evaluated in terms of speedup and energy against a server-grade multi-threaded CPU baseline. The results show that most PIM implementations can achieve higher throughput for different read lengths and edit distance thresholds. The state-of-the-art algorithm, WFA-adaptive, has up to 2-3 times speedup for all datasets, even when aligning large reads of lengths 5Kbp and 10Kbp, and achieves even higher speedup when the CPU-DPU communication time is not included.
