MVAPICH2

MVAPICH2 is a high-performance Message Passing Interface (MPI) library designed for cluster computing over high-speed interconnects such as InfiniBand, RoCE, iWARP, and Omni-Path. It provides robust MPI-1, MPI-2, and many MPI-3 features with an emphasis on low latency and high bandwidth for scalable applications. The project originated at The Ohio State University in collaboration with Mellanox Technologies and is the successor to the original MVAPICH project. It is distributed as open-source software under a BSD-style license and is widely deployed on Linux clusters in academic and research contexts.

Architecture and features: MVAPICH2 uses a transport-aware communication substrate that supports InfiniBand verbs and other network APIs, with a modular CH4-based path and, historically, the CH3 backend. It implements RDMA-based point-to-point and collective operations, supports GPU-aware communication via MVAPICH2-GDR for CUDA-enabled systems, and offers features such as multi-rail traffic, asynchronous progress, and optional fault-tolerance extensions in some variants. It aims to maximize data throughput and minimize CPU overhead by offloading communication to network hardware when possible.

Platform support and usage: MVAPICH2 runs on Linux and supports x86 and a variety of other architectures. It is commonly packaged with distribution tools and loaded via environment modules; users compile MPI applications with mpicc and run with mpirun or mpiexec. It also includes variants tailored to specific networks and is used in many national labs and HPC centers for large-scale simulations, data analysis, and scientific workloads.