Exploiting Content Similarity to Improve Memory Performance in Large-Scale High-Performance Computing Systems

Abstract

As we consider building the next generation of extreme-scale systems, many of the biggest challenges are related to memory characteristics. In particular, overcoming challenges related to resilience and memory bandwidth will require innovative strategies for improving the performance of main memory. In this paper, we propose to exploit memory content similarity to improve memory performance. We begin by presenting several novel strategies that leverage memory content similarity to improve system resilience and effective memory bandwidth. Additionally, we seek to understand the source of similarity in the memory of HPC applications.