Decentralized Memory Disaggregation Over Low-Latency Networks

Mosharaf Chowdhury is an Assistant Professor in the EECS Department at the University of Michigan. His research ranges from resource disaggregation in low-latency RDMA networks to geo-distributed analytics over the WAN, with a common theme of enabling applicationinfrastructure symbiosis across different layers of corresponding software and hardware stacks. [email protected] Memory disaggregation can expose remote memory across a cluster to local applications. However, existing proposals call for new architectures and/or new programming models, making them infeasible. We have developed a practical memory disaggregation solution, Infiniswap, which is a remote memory paging system for clusters with lowlatency, kernel-bypass networks such as RDMA. Infiniswap opportunistically harvests and transparently exposes unused memory across the cluster to unmodified applications by dividing the swap space of each machine into many chunks and distributing them to unused memory of many remote machines. For scalability, it leverages the power of many choices to perform decentralized memory chunk placements and evictions. Applications using Infiniswap receive large performance boosts when their working sets are larger than their physical memory allocations.