CosMoS: Architectural Support for Cost-Effective Data Movement in a Disaggregated Memory Systems

Memory disaggregation has emerged as a strong alternative to traditional server systems for improved memory utilization and scalability. The compute nodes with a small local memory are connected to disaggregated memory pools through memory-semantic interconnects, such as CXL, that support high-bandwidth and low-latency memory access. A primary concern with these systems is high remote memory latency due to the presence of a network interconnect between CPU and memory. A hot page migration is generally deployed in a hybrid memory system to improve the locality of memory access. In this article, we identify the major challenges in the present OS-based or hardware-based techniques for page migration in disaggregated memory. To this end, we propose CosMoS , an architectural solution for Cos t-effective data Mo vement in a S calable disaggregated memory system that can predict, schedule, and optimize the movement of hot pages between local and remote memory. Our mechanism also eliminates long delays on the critical demand memory accesses to other remote pages that are obstructed during hot page movement. We evaluate CosMoS over many data-centric workloads. Our results show a 20% performance improvement with CosMoS compared to the state-of-the-art mechanism and an 86% improvement compared to the baseline for a large-scale disaggregated memory system.