This article was accepted into the corpus but its outbound wikilinks were never NER-processed — typical at the deepest BFS hop or when the run's entity cap was reached. No expansion funnel to show.
| Cray MPI | |
|---|---|
| Name | Cray MPI |
| Developer | Cray Inc. |
| Released | 1990s |
| Latest release | (see vendor) |
| Operating system | Cray Linux Environment, SUSE Linux Enterprise Server |
| Platform | Cray systems, HPC clusters |
| License | Proprietary |
Cray MPI
Cray MPI is a proprietary implementation of the Message Passing Interface designed for high-performance computing systems produced by Cray Inc. and successor organizations. It targets tightly coupled supercomputers and large-scale clusters, emphasizing low-latency communication, scalable collective operations, and integration with vendor hardware. The implementation is widely used in scientific computing centers, national laboratories, and academic research facilities.
Cray MPI implements the MPI standard on architectures found in systems from Cray, enabling parallel applications developed for platforms such as HPE Cray EX Series, Cray XC50, and related supercomputers. It focuses on interoperability with job managers like SLURM, system software such as SUSE Linux Enterprise Server and the Cray Linux Environment, and middleware stacks common at facilities like Lawrence Livermore National Laboratory, Oak Ridge National Laboratory, and Argonne National Laboratory. Cray MPI provides bindings for languages used in scientific projects, including Fortran, C++, and Python (via MPI wrappers), and is tailored to leverage interconnects like Aries (interconnect) and Slingshot (network).
The implementation is engineered around Cray hardware topologies and fabric technologies. It integrates with hardware features such as the Cray XC series network topology, using low-level protocols optimized for the interconnect controller and NICs. Cray MPI employs a multi-tiered design that interfaces with system firmware, the kernel of SUSE Linux Enterprise Server, and runtime services provided by the resource manager (for instance, PBS Professional or SLURM). Process placement and mapping exploit knowledge of nodes and cabinets as in installations at National Center for Supercomputing Applications to minimize contention. The stack includes tuned collective algorithms, RDMA support for zero-copy transfers, and NUMA-aware optimizations tailored for processors from vendors like Intel and AMD and accelerator cards from NVIDIA.
Cray MPI extends the MPI standard with vendor-specific performance features and administrative controls. It provides enhanced collective operations, topology-aware communicators, and advanced point-to-point tuning for applications from communities such as climate modeling and computational chemistry centered at NOAA and Argonne National Laboratory. Fault-handling enhancements interoperate with system services used at Los Alamos National Laboratory and include diagnostics compatible with tooling like TAU Performance System and Intel VTune. Integration points for accelerators enable interoperability with programming models from OpenACC and CUDA used by researchers at institutions such as NASA and European Centre for Medium-Range Weather Forecasts. Administrative features support security and accounting frameworks common at national facilities, integrating with identity providers such as LDAP and site-specific schedulers.
Optimizations enable strong scaling on thousands of ranks and are validated on systems serving projects funded by agencies like the Department of Energy and National Science Foundation. Benchmarks at centers like Oak Ridge National Laboratory and Lawrence Livermore National Laboratory show low-latency latency for short messages and high bandwidth for large payloads when mapped to fabrics such as Aries (interconnect) and Slingshot (network). Algorithms for collectives draw from research communities exemplified by conferences such as SC (conference) and International Conference for High Performance Computing, Networking, Storage and Analysis. Performance tools from vendors and labs—CrayPAT, PAPI, and Perftools—are commonly used to profile MPI workloads and guide tuning across architectures including manycore systems from Intel and GPU-accelerated nodes using NVIDIA hardware.
Cray MPI is packaged to integrate with system images and application stacks used at national labs and universities, supporting compiler toolchains from GNU Project, Intel, and Cray Compiler Collection. It interoperates with software ecosystems including numerical libraries like LAPACK, ScaLAPACK, and PETSc, and domain codes such as GROMACS, LAMMPS, and WRF. The implementation adheres to MPI ABI conventions to facilitate use with container technologies championed by projects such as Singularity and Shifter, and it is often deployed alongside workflow managers used at research centers like Pegasus Workflow Management System and batch systems like Torque.
Development traces to Cray Research initiatives in the 1990s and continued through corporate transitions involving SGI and Hewlett Packard Enterprise. Incremental releases aligned with hardware generations such as the Cray XT, Cray XC, and the HPE Cray EX platforms, with collaboration between vendors and national laboratories to validate scalability for flagship systems at Argonne National Laboratory and Oak Ridge National Laboratory. Features and tuning evolved in response to research presented at venues like SC (conference) and initiatives sponsored by agencies including the Department of Energy. Open-source MPI implementations such as Open MPI and MPICH influenced standard compliance and interoperability testing, while vendor-specific extensions addressed facility needs.
Administrators deploy Cray MPI via system images and site-specific modules maintained at centers like National Energy Research Scientific Computing Center and Texas Advanced Computing Center. Users run scientific workloads in domains spanning geoscience, astrophysics, and bioinformatics on production queues managed by schedulers such as SLURM and PBS Professional. Best practices include binding processes to hardware affinity domains, leveraging topology-aware launchers, and profiling with tools used at facilities like Lawrence Berkeley National Laboratory. Training and documentation are typically provided by vendor support teams and through collaborations with institutions including University of Illinois Urbana–Champaign and University of California, Berkeley.
Category:Message Passing Interface Category:Supercomputing software