Title:Systems for Memory Disaggregation: Challenges & Opportunities

Abstract:Memory disaggregation addresses memory imbalance in a cluster by decoupling CPU and memory allocations of applications while also increasing the effective memory capacity for (memory-intensive) applications beyond the local memory limit imposed by traditional fixed-capacity servers. As the network speeds in the tightly-knit environments like modern datacenters inch closer to the DRAM speeds, there has been a recent proliferation of work in this space ranging from software solutions that pool memory of traditional servers for the shared use of the cluster to systems targeting the memory disaggregation in the hardware. In this report, we look at some of these recent memory disaggregation systems and study the important factors that guide their design, such as the interface through which the memory is exposed to the application, their runtime design and relevant optimizations to retain the near-native application performance, various approaches they employ in managing cluster memory to maximize utilization, etc. and we analyze the associated trade-offs. We conclude with a discussion on some open questions and potential future directions that can render disaggregation more amenable for adoption.