Understanding the Performance Impact of Queue-Based Resource Allocation in Scalable Disaggregated Memory Systems

Disaggregated memory systems are expected to replace traditional servers in next-generation data centers. Disaggregation overcomes the problem of memory under-utilization and scalability by allowing on-demand memory allocation to compute nodes from in-network memory pools. However, accessing on-network memory resources incurs additional access costs and significantly impacts the system’s performance, which is further impacted by multiple compute nodes sharing the same memory pools. However, the performance can be improved by ensuring fairness in queue allocation for network and remote memory resources to compute nodes running workloads with different access patterns. This paper explores the performance evaluation of multi-node disaggregated systems with different queue allocation methods for network and memory bandwidth partition. We propose using an in-network global memory controller to control the flow of memory requests to the remote memory pools to enforce Quality-of-Service. Finally, we utilize the memory request rate of each compute node to set their weights/priorities for different queue allocation methods. We evaluate all the scheduling policies using a custom trace-based disaggregated simulator over various benchmarks with different access patterns. Our results show that scheduling policies significantly impact the average memory latency and the system performance in different configurations.