Rama K. Govindaraju

Breaking the connection: RDMA deconstructed

The architecture, design and performance of RDMA (remote direct memory access) over the IBM HPS (high performance switch and adapter) are described. Unlike conventional implementations such as InfiniBand, our RDMA transport model is layered on top of an unreliable datagram interface, while leaving the task of enforcing reliability to the ULP (upper layer protocol). We demonstrate that our model allows a single MPI task to deliver bidirectional bandwidth of close to 3.0 GB/s across a single link and 24.0 GB/s when striped across 8 links. In addition, we show that this transport protocol has superior attributes in terms of a) being able to handle RDMA packets coming out of order; b) being able to use multiple routes between a source-destination pair and c) reducing the size of adapter caches.

Architecture and early performance of the new IBM HPS fabric and adapter

In this paper we describe the architecture, design, and performance of the new cluster switch fabric and adapter called HPS (High Performance Switch). HPS delivers very low latency and very high bandwidth. We demonstrate latency of less than 4.3us MPI library; 1.8GB/s of delivered unidirectional bandwidth and 2.9GB/s of bidirectional bandwidth between 2 MPI tasks running on 1.9GHz Power 4+ IH based nodes. HPS also supports RDMA (remote direct memory access capability). A unique capability of RDMA over HPS is that reliable RDMA is supported over an underlying unreliable transport (unlike Infiniband and other RDMA transport protocols which depend on the underlying transport being reliable). We profile the performance of RDMA and its impact on striping for systems in which multiple network adapters are available to tasks of parallel jobs.