MPI MiS

MPI MiS
Appaloosa · CC BY-SA 3.0 · source
Name	MPI MiS
Developer	Max Planck Society
Released	20XX
Programming language	C (programming language), Fortran
Operating system	Linux, Unix
License	Proprietary / Academic

MPI MiS

MPI MiS is a parallel computing middleware and message-passing implementation developed to support high-performance scientific computing on distributed systems. It combines low-latency communication primitives with resource management features to serve workloads in computational physics, climate modeling, and large-scale data analysis. The project interfaces with established numerical libraries, scheduling systems, and hardware vendors to integrate into research infrastructures such as national supercomputing centers and university clusters.

Overview

MPI MiS provides a message-passing interface that interoperates with implementations and standards from projects such as Message Passing Interface, Open MPI, MPICH, Intel MPI Library, and MVAPICH. It targets hardware platforms from vendors including Intel Corporation, Advanced Micro Devices, NVIDIA Corporation, and IBM by supporting network fabrics like InfiniBand, Omni-Path, and Ethernet (computer networking). MPI MiS is designed to work with parallel programming models exemplified by OpenMP, CUDA, OpenCL, and Partitioned Global Address Space adaptations, and to be integrated into workflow systems such as Slurm Workload Manager, PBS (software), and HTCondor.

History and Development

Development of MPI MiS traces to collaborative efforts among research institutes and centers such as the Max Planck Institute for Informatics, Lawrence Berkeley National Laboratory, Argonne National Laboratory, Los Alamos National Laboratory, and European partners like CERN. Early design decisions were influenced by findings from projects including Blue Gene architectures, the TOP500 ranking community, and initiatives at Deutsches Elektronen-Synchrotron to optimize interconnect performance. Contributors included teams with backgrounds linked to organizations such as Cray Research, Fujitsu, and the European Centre for Medium-Range Weather Forecasts where scalable communication libraries were critical. Funding and oversight involved national research councils such as the Deutsche Forschungsgemeinschaft, European Research Council, and government agencies like the U.S. Department of Energy.

Architecture and Design

The architecture employs a modular design with layers comparable to designs in OpenMPI and MPICH, separating transport, collective operations, and point-to-point messaging. Key components mirror patterns established by ZeroMQ and gRPC for message framing, while collective algorithms draw on research associated with ScaLAPACK, PETSc, Trilinos, and HPCX. The transport abstraction supports offload engines similar to RDMA mechanisms on InfiniBand and leverages NIC capabilities from vendors like Mellanox Technologies. Fault tolerance mechanisms take inspiration from checkpoint/restart frameworks such as BLCR and Checkpoint/Restart in HPC, and are tailored for job managers including SLURM and Grid Engine (software).

Implementation and Deployment

Implementations of MPI MiS are provided as compiled libraries and language bindings for C (programming language), C++, and Fortran, with interoperability layers for Python (programming language) via bridges to NumPy and SciPy. Deployment practices follow conventions from supercomputing centers such as Jülich Research Centre, Oak Ridge National Laboratory, Riken, and National Energy Research Scientific Computing Center, including containerized delivery through Docker (software) and Singularity (software). Integrations with job schedulers like Slurm Workload Manager and PBS Professional facilitate resource allocation across partitions in facilities such as HPC Wales and regional grids coordinated by European Grid Infrastructure.

Performance and Benchmarks

Performance characterization references methodologies from the HPCG benchmark and LINPACK-style testing used by the TOP500 community, while microbenchmarks borrow from suites like OSU Micro-Benchmarks and Intel MPI Benchmarks. Results reported by centers including Forschungszentrum Jülich, Argonne National Laboratory, and Lawrence Livermore National Laboratory demonstrate low-latency point-to-point messaging competitive with Open MPI and MPICH on fabrics provided by Mellanox Technologies and Intel Corporation. Collective operation optimizations build on algorithms pioneered in HPC libraries and research from universities such as ETH Zurich, MIT, Stanford University, and University of California, Berkeley.

Applications and Use Cases

MPI MiS is applied in domains that depend on scalable message passing, including climate simulation projects at Max Planck Institute for Meteorology, astrophysics codes developed at European Southern Observatory, computational chemistry packages associated with Max Planck Institute for Coal Research, and materials science efforts at Argonne National Laboratory. It is used alongside software frameworks like WRF (model), GROMACS, LAMMPS, NAMD, Quantum ESPRESSO, and VASP to accelerate production workflows in research institutions such as University of Cambridge, Harvard University, Princeton University, and University of Oxford.

Security and Compliance

Security features align with best practices in HPC centers overseen by organizations such as National Institute of Standards and Technology, European Union Agency for Cybersecurity, and compliance regimes that reference standards like ISO/IEC 27001 when relevant. Network-level protections integrate with authentication and authorization systems common to research infrastructures, including federations like eduGAIN and identity providers used by PRACE and national labs. Vulnerability handling follows disclosure processes used by vendors such as Red Hat and advisory channels maintained by entities like CERT Coordination Center.

Category:Parallel computing software