Platform MPI has three strategies to optimize performance on multi-core systems:

• Direct shared memory implementation of MPI collective operations (SHMEM)
• Policy based process-to-core affinity binding (including support for LSF syntax)
• Multi-core-aware copying and encoders/decoders optimized for inter-core memory hierarchy

Platform MPI has an extensive set of tools including verification and test tools, application tracing and timing facilities, and performance counters with message sizes and counts. These tools enable application developers and users to maximize application performance.

MPI-related monitoring presentation can be selected through environment variables to determine presentation of timing and trace information, with no recompilation or re-linking of the application. There is no need for re-compilation or re-linking the application.

Multithreaded applications can fully exploit Platform MPI and multiple threads can simultaneously request services and conduct communication.

Through the use of a priority network list built from system configuration files, user environment variables, user command line options and library hard-coded defaults, Platform MPI can dynamically select the optimal network connection between each node and each other node within a cluster at runtime. This maximizes network efficiency so that processes get the data they need quickly.

Platform MPI enables you to use a single MPI implementation, and a single executable, to support a wide array of interconnects, with dynamic runtime selection of interconnect. With Platform MPI, you can run a single application across multiple architectures, Linux distributions and interconnects. It eliminates need to recompile applications and MPI libraries, while providing full support for dynamically linked libraries. These means there is no need to recompile for every compiler/interconnect combination.

Command line arguments to the application are automatically provided to all MPI processes, avoiding tedious parsing and broadcasting of parameters to other MPI processes.

The Multiple Instruction – Multiple Data (MIMD) model is supported through provisions that launch different executables which constitute the whole MPI application.