RDMA: Provably More Powerful Communication

Опубликовано 21 октября 2019, 16:23

In recent years, a technology known as Remote Direct Memory Access (RDMA) has made its way into data centers, earning a spotlight in distributed systems research. RDMA provides the traditional send/receive communication primitives, but also allows a process to directly read/write remote memory. Recent work shows that RDMA leads to some new and exciting distributed algorithms. However, it is unclear whether the performance benefits observed with RDMA are inherent to its non-traditional features, or just an artifact of the way it is currently being used.

In this talk, I’ll present a theoretical model of RDMA, called the Message-and-Memory model, and use it to show that RDMA has some fundamental advantages over previous communication mechanisms. I’ll use the classic problem of consensus to analyze the benefits that RDMA brings, and prove upper and lower bounds on fault tolerance and performance of solving crash-fault-tolerant and Byzantine-fault-tolerant consensus in this model.

See more at microsoft.com/en-us/research/v...