Open MPI

Open MPI
Open MPI · BSD · source
Name	Open MPI
Developer	Open MPI Project
Initial release	2004
Programming language	C, C++, Fortran
Operating system	Unix-like, Linux, macOS
License	BSD-style

Open MPI is an open-source Message Passing Interface implementation that provides a high-performance, portable, and scalable runtime for parallel applications. It integrates contributions from multiple research institutions and vendors to support distributed-memory computing across clusters, supercomputers, and cloud platforms. Open MPI is widely used in scientific computing, engineering, and data-intensive research projects.

Overview

Open MPI is designed to support applications developed with the Message Passing Interface standard and to interoperate with a broad range of hardware and software ecosystems. It targets high-performance computing centers such as Argonne National Laboratory, Oak Ridge National Laboratory, Lawrence Berkeley National Laboratory, Los Alamos National Laboratory, and Sandia National Laboratories, while also serving academic groups at Massachusetts Institute of Technology, Stanford University, University of California, Berkeley, University of Illinois Urbana-Champaign, and Georgia Institute of Technology. The project collaborates with vendors including IBM, Intel, NVIDIA, Cray Inc., and Hewlett Packard Enterprise to optimize for interconnects like InfiniBand, Omni-Path, Mellanox Technologies, Intel Omni-Path Architecture, and Ethernet. Open MPI supports programming languages used in computational science communities, including Fortran, C++, and Python via bindings and wrappers supported by ecosystems like NumPy, SciPy, PETSc, and Trilinos.

History and Development

Open MPI originated from a consortium formed in response to fragmentation in MPI implementations, with founding contributors from Los Alamos National Laboratory, Argonne National Laboratory, University of New Mexico, IBM, and Intel. The project emerged during a period of parallel runtime innovation alongside contemporaries such as MPICH, LAM/MPI, and vendor stacks from HP and Cray Research. Key development milestones involved integration of collective algorithms influenced by research from Lawrence Livermore National Laboratory and scheduling strategies used in systems like Blue Gene and Titan. Governance and roadmap decisions have been influenced by collaborations with centers including National Energy Research Scientific Computing Center, Oak Ridge Leadership Computing Facility, and international partners such as Cineca and EPCC.

Architecture and Components

Open MPI's modular architecture consists of layered subsystems that communicate through well-defined frameworks. Core components include the runtime environment derived from work on process management at Sun Microsystems and UNIX System Laboratories, point-to-point messaging layers comparable to designs in MPICH, collective communication modules influenced by algorithms from MPI Forum contributions, and resource managers that integrate with schedulers like Slurm Workload Manager, TORQUE, PBS Professional, and LSF. Network transport plugins interface with low-level drivers from Mellanox Technologies, Intel, Broadcom, and Cisco Systems. Process launch and control subsystems draw on standards used by Open Container Initiative and concepts from POSIX process control. The project also provides interfaces to hardware locality tools such as hwloc and interoperability with debuggers and profilers like TotalView, Allinea DDT, and Valgrind.

Features and Implementation Details

Open MPI implements MPI point-to-point and collective operations with tunable algorithms for topologies including Cartesian and graph mappings defined by the MPI Forum standard. It supports advanced MPI features like one-sided communication, dynamic process management, and parallel I/O that builds on concepts from MPI-IO and integrates with storage solutions at centers such as NERSC and Fermilab. Implementation details include modular BTL (Byte Transfer Layer), PML (Point-to-Point Messaging Layer), and MTL (Matching Transport Layer) plugins, which allow selection among transports like InfiniBand, shared memory, and TCP/IP stacks used by Red Hat and SUSE. Fault tolerance research contributions tie to projects at University of Edinburgh and Technische Universität München, while collective tuning leverages autotuning frameworks from Argonne National Laboratory and empirical studies published in venues such as SC Conference and IEEE International Parallel and Distributed Processing Symposium.

Performance and Scalability

Performance tuning in Open MPI addresses latency, bandwidth, and scalability across node counts ranging from small clusters to leadership systems like Summit, Fugaku, and Perlmutter. Benchmarks compare Open MPI to implementations such as MPICH and vendor MPI stacks from Cray Inc. and HPE using suites like OSU Micro-Benchmarks, Intel MPI Benchmarks, and application proxies from LINPACK and SPEC. Optimizations exploit hardware features from ARM Holdings architectures used in some supercomputers, vectorization from Intel Xeon and AMD EPYC processors, and accelerators such as NVIDIA Tesla and AMD Radeon Instinct. Scalability studies reference machine characteristics reported by TOP500 and software trace analyses from tools like TAU Performance System and Scalasca.

Ecosystem and Implementations

Open MPI is part of a wider ecosystem including numerical libraries and frameworks such as PETSc, Trilinos, HDF5, NetCDF, Boost, and Eigen. Containerized deployments leverage standards from Docker, Singularity, and Kubernetes orchestrated on cloud providers like Amazon Web Services, Google Cloud Platform, and Microsoft Azure. Integration with build and package systems includes CMake, Autoconf, RPM Package Manager, and Debian packaging workflows. Competitive and complementary MPI implementations include MPICH, Intel MPI Library, MVAPICH2, and legacy systems like LAM/MPI.

Adoption and Use Cases

Open MPI is used extensively in domains such as climate modeling at centers like NOAA, computational chemistry in collaborations with Argonne National Laboratory and Lawrence Berkeley National Laboratory, astrophysics groups at Caltech and University of Cambridge, and bioinformatics workflows in projects at European Bioinformatics Institute and Broad Institute. Industrial adopters include Schlumberger, ExxonMobil, and Siemens for simulation and optimization workloads. Educational courses at institutions such as Massachusetts Institute of Technology, Princeton University, and ETH Zurich teach parallel programming using MPI examples built on Open MPI. Research publications citing Open MPI appear in journals like Journal of Parallel and Distributed Computing, ACM Transactions on Mathematical Software, and proceedings from SC Conference and IEEE Cluster Conference.

Category:Message Passing Interface