Scatterv

Scatterv is a collective communication operation in the Message Passing Interface (MPI) that distributes data from a root process to all processes in a communicator, allowing each recipient to receive a different amount of data. It generalizes MPI_Scatter by permitting variable block sizes, enabling nonuniform data distribution across processes.

In MPI, the C signature is int MPI_Scatterv(const void *sendbuf, const int *sendcounts, const int *displs, MPI_Datatype sendtype, void *recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm). Other language bindings have analogous forms. Key parameters are:

- sendbuf: starting address of the data to send (significant only at the root process).

- sendcounts: array of length equal to the number of processes, specifying how many elements of sendtype to send to each process.

- displs: array of length equal to the number of processes, specifying the starting offset in sendbuf for each recipient. Offsets are in units of sendtype.

- sendtype: datatype of the data being sent.

- recvbuf: buffer in the receiving process to receive the data.

- recvcount: maximum number of elements that can be received by a process; all processes provide this value, and the actual number received by a given process is sendcounts[rank].

- recvtype: datatype of the data being received.

- root: rank of the root process that provides the data.

- comm: communicator over which the operation is performed.

Non-root processes ignore sendbuf, sendcounts, and displs; they must supply a recvbuf large enough for at least recvcount elements and receive recvcount elements of recvtype.

The sum of all sendcounts must equal the total number of elements in the send buffer. Displacements and counts are defined in units of sendtype, and derived datatypes are permitted.

Example pattern: with N processes, root 0 has a sendbuf of total M elements, and sendcounts may be [3, 2, 4, 1] with displs [0, 3, 5, 9] for M = 10; each process receives a different portion of the data accordingly, with the maximum per-process receive limit set by recvcount. Scatterv is useful when data ownership or workload varies across processes.