Remote regions: a simple abstraction for remote memory

Abstract

We propose an intuitive abstraction for a process to export its memory to remote hosts, and to access the memory exported by others. This abstraction provides a simpler interface to RDMA and other remote memory technologies compared to the existing verbs interface. The key idea is that a process can export parts of its memory as files, called remote regions, that can be accessed through the usual file system operations (read, write, memory map, etc). We built this abstraction in the Linux kernel, and evaluated it. We show that remote regions are easy to use and perform close to RDMA. We demonstrate it via micro-benchmarks and by adapting two in-memory single-host applications to use remote memory: R and Metis. With R, using remote regions requires no changes to the code and allows R to run with remote memory that exceeds the physical memory of a host. With Metis, the modifications amount to ≈100 lines of code and they allow Metis to scale its performance across 8 hosts.