MPIScatter

MPIScatter is a collective communication operation in the Message Passing Interface (MPI) that distributes blocks of data from one root process to all processes within a communicator. Each process receives a distinct block of data, enabling data parallelism across the participating processes.

In a typical scatter operation, the root process provides a send buffer containing a sequence of data elements arranged for all processes. The data is divided into equal-sized blocks, with sendcount elements of type sendtype allocated to each process. The process with rank i in the communicator receives its block into its own receive buffer, recvbuf, which holds recvcount elements of type recvtype. All processes in the communicator participate in the call; the root process is responsible for supplying the send buffer, while non-root processes supply a receive buffer. The contents of the send buffer are mapped to processes in rank order.

The standard prototype in C is MPI_Scatter(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, root, comm). In Fortran, the order and types differ, but the conceptual behavior remains the same. On non-root processes, the sendbuf parameter is ignored; on the root, the sendbuf must contain at least sendcount multiplied by the number of processes elements, arranged so that each process receives a contiguous block.

MPIScatter is often used as a building block for parallel algorithms that require dividing input data evenly among processes, such as distributed initializations, processing of independent data chunks, or parallel reductions following per-process computations. Related operations include MPI_Scatterv, which allows varying block sizes, and MPI_Gather, which reverses the data distribution.