MPIs

MPIs refer to the implementations of the Message Passing Interface (MPI) standard, the de facto API for programming distributed-memory parallel computers. The MPI standard, maintained by the MPI Forum, specifies a portable set of routines that enable processes to communicate and synchronize across hardware and networks.

The MPI model comprises parallel processes that communicate through communicators, ranks, and tags. It provides point-to-point communication via MPI_Send and MPI_Recv, and a rich set of collective operations such as MPI_Bcast, MPI_Scatter, MPI_Gather, MPI_Reduce, and MPI_Alltoall. Non-blocking variants (MPI_Isend, MPI_Irecv) and persistent calls support overlap of computation and communication. Derived data types, groups, and topologies help optimize data handling and process layout.

Prominent MPIs include MPICH, Open MPI, MVAPICH, and Intel MPI. These implementations aim for standards conformance and interoperability, and run on a wide range of networks and hardware.

Applications are typically written in C, C++, or Fortran and follow the single program, multiple data model. Programs call MPI_Init at start and MPI_Finalize at end, determine their rank and the communicator size with MPI_Comm_rank and MPI_Comm_size, and perform communication accordingly.

MPI has evolved through major revisions (MPI-1, MPI-2, MPI-3, and MPI-4), adding features such as dynamic process management, enhanced I/O, neighborhood collectives, and improved one-sided communication.

In practice, MPIs underpin much of high-performance computing on clusters and supercomputers. They offer portability across platforms but introduce complexity and overhead; performance depends on library quality, network characteristics, and program design, and many deployments combine MPI with shared-memory models like OpenMP.