Open Collections

Ramezanikebrya, Hamidreza

University of British Columbia

Processing-in-Memory (PiM) has gained attention as a potential solution to address the "memory wall" in high-performance computing applications. Recently the first general-purpose PiM technology - dubbed UPMEM - has been made commercially available. While several studies report substantial advantage (often as application speedup) of UPMEM-equipped servers over high-end CPUs, these findings are often "fragile" due to overlooked factors, such as: (i) using unoptimized code on either one or both sides of the comparison, (ii) excluding overheads inherent to the accelerator-like processing model offered by UPMEM, (iii) ignoring power or energy consumption in the comparison, and (iv) high dependence on the the choice of CPU hardware baseline which can dramatically change the conclusions. This thesis aims to address these limitations by reexamining UPMEM PiM’s effectiveness within specific application contexts. To that end, we focus on applications that are well-suited for UPMEM, primarily in the area of genomic sequence alignment, and conduct a rigorous performance evaluation, ensuring both UPMEM and CPU codes are reasonably optimized, extending the comparison to include power consumption, and performing cost- and power-normalized comparisons to mitigate hardware dependency. Our experiments reveal that, after applying optimizations to CPU baselines, the previously observed performance gap narrows significantly, and in many cases, the CPU-equipped servers offer an advantage even in terms of raw performance. More importantly, cost- and power-normalized comparisons suggest that while UPMEM may offer potential benefits in the future, the current technology is unlikely to offer a cost- or power-normalized advantage compared to CPU baselines for the set of applications we evaluate. Finally, we provide practical heuristics to guide decisions on which types of computational tasks are best suited for offloading to UPMEM PiM system. These insights aim to assist researchers and practitioners in effectively leveraging UPMEM technology where may offer an advantage.

Text

2024-12-02

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/89771

Electrical and Computer Engineering

University of British Columbia

UBCV

http://creativecommons.org/licenses/by-nc-nd/4.0/