Towards Embracing Object Granularity in Tiered Memory Management for Big Data
Maciej Kokociński, Tadeusz Kobus, Krystian Chmielewski, Rafał Pyzik, Maciej Maciejewski · 2024
The evolving landscape of computer memory architectures necessitates innovative approaches to memory management. This paper explores the potential of object-based memory management in tiered memory systems. Our study of HiBench workloads reveals significant variance in object access frequency and asymmetry in read/write operations. Using a simulation of a two-tier memory architecture, we demonstrate that optimal object placement can reduce accesses to the slow memory tier by 60% compared to traditional page migration techniques. Additionally, we find that static code analysis is insufficient for predicting object hotness, underscoring the need for dynamic memory management strategies.