Open Collections

Gharavi, Niloofar

University of British Columbia

In modern computing systems, the memory wall challenge poses a significant obstacle to performance optimization. This challenge refers to the widening gap between CPU processing speeds and memory access speeds, which has become increasingly pronounced with advancements in CPU technology. As CPUs continue to accelerate in performance, the pace of improvement in memory access speeds has not kept up, leading to a bottleneck in data-intensive applications. Processing-in-Memory (PIM) emerges as a promising solution by bridging the gap between computation and data. However, the current architectural limitations, such as an incomplete set of instructions, and inter-processor communication, pose challenges in identifying workloads that yield significant performance benefits. In this thesis, we investigate four applications—image decoding, key-value lookups, data filtration, and text search—to discern the potential impact of PIM hardware. Our study reveals a spectrum of outcomes, ranging from best-case scenarios to worst-case scenarios for PIM-accelerated workloads. While the worst-case analysis demonstrates a potential increase in power consumption without a corresponding improvement in system throughput, the best-case scenarios illustrate a remarkable performance boost of up to 14 times. This research contributes valuable insights into the application of PIM for specific workloads, which helps better understand its potential benefits and limitations in varied data processing applications.

Text

2024-03-27

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/87638

Electrical and Computer Engineering

University of British Columbia

UBCV

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