Huge page aware memory management scheme for disaggregated memory

Abstract

Nowadays, emerging technologies such as big data analysis and machine learning impose a significant burden on memory costs. To address these problems, the concept of a disaggregated memory system is gaining traction as a means to efficiently utilize memory resources. Page migration becomes necessary to achieve high memory utilization in a disaggregated memory system. Furthermore, employing huge pages can enhance the overall performance of applications on a disaggregated memory system by reducing address translation costs. However, the use of huge pages can also result in increased memory internal fragmentation, leading to wasted local memory. To mitigate this issue, we propose a novel page migration technique that incorporates effective huge page promotion. Our scheme demonstrates a 2.3% performance improvement compared to a scenario where all memory resides in local memory, despite our scheme utilizing only 30% of local memory. Additionally, our approach effectively leverages huge pages, reducing TLB misses by 75% compared to a scheme that only employs regular page sizes.