Exploring the Disaggregated Memory Interface Design Space

Abstract: Rapid increases in network performance, coupled with the dynamic nature of modern cloud workloads, has led to calls for memory disaggregation within the datacenter. In a memory disaggregated machine, compute nodes contain only a limited amount of local memory, but are able to utilize free memory elsewhere in the machine over a network connection. This broad definition of memory disaggregation allows for a great diversity of techniques to actually provide that interface. In this paper, we present an overview of these techniques and lay out a few dimensions of the disaggregated memory interface design space. We conclude with a discussion of future research directions and challenges.