Disaggregated Memory Benefits for Server Consolidation

Recent architecture research has introduced a new building block - a memory blade - which provides disaggregated memory capacity expansion and sharing for an ensemble of blade servers. In this paper, we examine the systems implications of this new architectural building block. We build a software-based prototype of memory disaggregation and examine how the additional level of indirection provided by the memory blade can provide significantly higher levels of consolidation. We specifically examine the use case of multiple large memory virtual machines (VMs) consolidated onto a single server. We explore content based sharing strategies to maximize the utilization of both local and remote memory. Using an in-memory database workload, our results show significantly higher levels of consolidation versus baseline servers (twice as many VMs than a memory-constrained baseline, providing 47% higher throughput), cost-effective memory expansion (28% better performance-per-dollar versus a large memory baseline), and effective content based sharing (up to 40%).