Amanda Baran - Distributed Computing with RDMA

Abstract : Remote Direct Memory Access technologies make it possible for distributed applications to interact with each others’ memory using one-sided verbs. This gives the programmer the illusion of a single global address space, and makes it possible to write distributed applications in a manner that is similar to how shared memory concurrent applications are written. In this talk, we will first provide a background knowledge in RDMA technologies, highlighting the ultra-low latency and high bandwidth promises of the one-sided model. However, the reality of RDMA programming reveals a disconnect between the theoretical promises of RDMA and its practical performance. To bridge this gap, this talk introduces a set of practical guidelines for high-performance RDMA programming, such as smart Queue Pair (QP) management and the minimization of control-path operations. To address one major challenge — the lack of guaranteed atomicity between local and remote memory accesses — we present ALock, an asymmetric lock primitive. ALock synchronizes local and remote processes using a combination of MCS queues and Peterson's algorithm, significantly outperforming traditional spinlocks and MCS locks under high contention. Finally, we demonstrate Remus, a framework designed to simplify the RDMA programming model. By moving away from complex, low-level verb code to an intuitive shared-memory API, Remus allows developers to perform operations like an RDMA read with a single line of code.