RMem: An OS Service for Transparent Remote Memory Access in Lightweight Manycores

Overview & Synopsis

Lightweight manycores deliver high performance and scal-ability at low power consumption. However, architectural intricacies of these processors impose programmability challenges that keep them away from mass adoption. While several efforts aim at introducing parallel programming environments to lightweight manycores, few initiatives are concerned about how to design rich Operating Systems (OSs) to them. In this work, we focus on the open challenges that arise from constrained memory subsystems of lightweight manycores, such as the presence of multiple address spaces and limited on-chip memory. To cope with transparent data access in this scenario, we introduce an OS service, named RMem. This service provides a shared memory abstraction over multiple address spaces and exposes system calls that enable one-sided communication on top of this abstraction. We implemented a prototype of our service in the Nanvix research OS, and we deployed the system the Kalray MPPA-256 lightweight manycore. Our experimental results with a microbenchmark unveiled that, while exposing an easier-to-program interface, the RMem Service may deliver about 91% of the write performance and up to 2.4× better read performance than the primitives in the libraries of the experimental platform.