MPI_REDUCE_SCATTER, MPI_Reduce_scatter

Purpose

Applies a reduction operation to the vector sendbuf over the set of tasks
specified by comm and scatters the result according to the values in
recvcounts.

C synopsis

#include <mpi.h>
int MPI_Reduce_scatter(void* sendbuf,void* recvbuf,int *recvcounts,
                       MPI_Datatype datatype,MPI_Op op,MPI_Comm comm);

C++ synopsis

#include mpi.h
void MPI::Comm::Reduce_scatter(const void* sendbuf, void* recvbuf,
                               int recvcounts[],
                               const MPI::Datatype& datatype,
                               const MPI::Op& op) const;

FORTRAN synopsis

include 'mpif.h' or use mpi
MPI_REDUCE_SCATTER(CHOICE SENDBUF,CHOICE RECVBUF,INTEGER RECVCOUNTS(*),
                   INTEGER DATATYPE,INTEGER OP,INTEGER COMM,INTEGER IERROR)

Description

This subroutine first performs an element-wise reduction on a vector of
count = the sum of i recvcounts[i] elements in the send buffer defined by
sendbuf, count, and datatype. Next, the resulting vector is split into n
disjoint segments, where n is the number of members in the group. Segment i
contains recvcounts[i] elements. The ith segment is sent to task i and
stored in the receive buffer defined by recvbuf, recvcounts[i], and
datatype.

MPI_REDUCE_SCATTER is functionally equivalent to MPI_REDUCE with count
equal to the sum of recvcounts[i] followed by MPI_SCATTERV with sendcounts
equal to recvcounts.

The "in place" option for intracommunicators is specified by passing
MPI_IN_PLACE in the sendbuf argument. In this case, the input data is taken
from the top of the receive buffer. The area occupied by the input data may
be either longer or shorter than the data filled by the output data.

If comm is an intercommunicator, the result of the reduction of the data
provided by tasks in group A is scattered among tasks in group B, and vice
versa. Within each group, all tasks provide the same recvcounts argument,
and the sum of the recvcounts entries should be the same for the two
groups.

MPI_IN_PLACE is not supported for intercommunicators.

When you use this subroutine in a threads application, make sure all
collective operations on a particular communicator occur in the same order
at each task. See IBM Parallel Environment for AIX: MPI Programming Guide
for more information on programming with MPI in a threads environment.

Parameters

sendbuf
     is the starting address of the send buffer (choice) (IN)
recvbuf
     is the starting address of the receive buffer (choice) (OUT)
recvcounts
     integer array specifying the number of elements in result distributed
     to each task. Must be identical on all calling tasks. (IN)
datatype
     is the datatype of elements in the input buffer (handle) (IN)
op
     is the reduction operation (handle) (IN)
comm
     is the communicator (handle) (IN)
IERROR
     is the FORTRAN return code. It is always the last argument.

Notes

See IBM Parallel Environment for AIX: MPI Programming Guide.

The MPI standard urges MPI implementations to use the same evaluation order
for reductions every time, even if this negatively affects performance. PE
MPI adjusts its reduce algorithms for the optimal performance on a given
task distribution. The MPI standard suggests, but does not mandate, this
sacrifice of performance. PE MPI chooses to put performance ahead of the
MPI standard's recommendation. This means that two runs with the same task
count may produce results that differ in the least significant bits, due to
rounding effects when evaluation order changes. Two runs that use the same
task count and the same distribution across nodes will always give
identical results.

In the 64-bit library, this function uses a shared memory optimization
among the tasks on a node. This optimization is discussed in the chapter
"Using shared memory" of IBM Parallel Environment for AIX: MPI Programming
Guide, and is enabled by default. This optimization is not available to
32-bit programs.

Errors

Fatal errors:

Invalid recvcounts
     recvcounts[i] < 0

Invalid datatype

Type not committed

Invalid op

Invalid communicator

Unequal message lengths

Invalid use of MPI_IN_PLACE

MPI not initialized

MPI already finalized

Develop mode error if:

Inconsistent op

Inconsistent datatype

Related information

     MPE_IREDUCE_SCATTER
     MPI_OP_CREATE
     MPI_REDUCE