Recent Advances in Overcoming Bottlenecks in Memory Systems and Managing Memory Resources in GPU Systems

TL;DR

This paper reviews recent research on memory system bottlenecks in GPU systems, focusing on contention, latency, and resource management to improve performance and efficiency.

Contribution

It provides an extended summary of innovative solutions addressing memory interference, latency, and resource management challenges in GPU systems.

Findings

Memory contention significantly impacts GPU performance.

New techniques reduce memory latency and overheads.

Improved memory management enhances energy efficiency.

Abstract

This article features extended summaries and retrospectives of some of the recent research done by our research group, SAFARI, on (1) various critical problems in memory systems and (2) how memory system bottlenecks affect graphics processing unit (GPU) systems. As more applications share a single system, operations from each application can contend with each other at various shared components. Such contention can slow down each application or thread of execution. The compound effect of contention, high memory latency and access overheads, as well as inefficient management of resources, greatly degrades performance, quality-of-service, and energy efficiency. The ten works featured in this issue study several aspects of (1) inter-application interference in multicore systems, heterogeneous systems, and GPUs; (2) the growing overheads and expenses associated with growing memory densities and latencies; and (3) performance, programmability, and portability issues in modern GPUs, especially those related to memory system resources.