Optimizing MPICH2 Performance on InfiniBand with RDMA

Title: Optimizing MPICH2 Performance on InfiniBand with RDMA

Abstract: This research aimed to design and implement an efficient version of MPICH2, a parallel processing standard, on InfiniBand with Remote Direct Memory Access (RDMA) support. The goal was to exploit RDMA operations and achieve high performance. The study focused on optimizing MPI-1 functions and proposed a zero-copy-based design. The impact of optimizations and designs was characterized using microbenchmarks. The application-level evaluation used the NAS Parallel Benchmarks. The optimized MPICH2 implementation achieved 7.6 µs latency and 857 MB/s bandwidth, close to the raw performance of the InfiniBand platform. The RDMA Channel interface in MPICH2 provided a simple, yet powerful, abstraction that enabled high-performance implementations with less development effort.

Link to Article: https://arxiv.org/abs/0310059v1 Authors: arXiv ID: 0310059v1