Tackling Diversity and Heterogeneity by Vertical Memory Management

TL;DR

This paper introduces a vertical memory management approach that reduces resource contention across multiple memory hierarchy levels, enhancing flexibility and efficiency for diverse workloads in heterogeneous memory systems.

Contribution

It proposes a novel vertical partitioning technique combined with horizontal policies to improve memory management in heterogeneous systems, addressing contention and workload diversity.

Findings

Reduces memory resource contention at multiple hierarchy levels

Supports flexible memory management policies for diverse workloads

Suitable for future heterogeneous memory architectures

Abstract

Existing memory management mechanisms used in commodity computing machines typically adopt hardware based address interleaving and OS directed random memory allocation to service generic application requests. These conventional memory management mechanisms are challenged by contention at multiple memory levels, a daunting variety of workload behaviors, and an increasingly complicated memory hierarchy. Our ISCA-41 paper proposes vertical partitioning to eliminate shared resource contention at multiple levels in the memory hierarchy. Combined with horizontal memory management policies, our framework supports a flexible policy space for tackling diverse application needs in production environment and is suitable for future heterogeneous memory systems.