1
1
openmpi/ompi/mpi/man/man3/MPI_Neighbor_alltoall.3in
Nathan Hjelm c5596548b2 MPI-3: Add support for neighborhood collectives
Blocking versions are simple linear algorithms implemented in coll/basic. Non-
blocking versions are from libnbc 1.1.1. All algorithms have been tested with
simple test cases.

cmr=v1.7.4:reviewer=jsquyres

This commit was SVN r29265.
2013-09-26 21:55:08 +00:00

152 строки
6.2 KiB
Plaintext

.\" Copyright 2013 Los Alamos National Security, LLC. All rights reserved.
.\" Copyright 2010 Cisco Systems, Inc. All rights reserved.
.\" Copyright 2006-2008 Sun Microsystems, Inc.
.\" Copyright (c) 1996 Thinking Machines Corporation
.TH MPI_Neighbor_alltoall 3 "#OMPI_DATE#" "#PACKAGE_VERSION#" "#PACKAGE_NAME#"
.SH NAME
\fBMPI_Neighbor_alltoall\fP \- All processes send data to neighboring processes in a virtual topology communicator
.SH SYNTAX
.ft R
.SH C Syntax
.nf
#include <mpi.h>
int MPI_Neighbor_alltoall(void *\fIsendbuf\fP, int \fIsendcount\fP,
MPI_Datatype \fIsendtype\fP, void *\fIrecvbuf\fP, int \fIrecvcount\fP,
MPI_Datatype \fIrecvtype\fP, MPI_Comm \fIcomm\fP)
.fi
.SH Fortran Syntax
.nf
INCLUDE 'mpif.h'
MPI_NEIGHBOR_ALLTOALL(\fISENDBUF, SENDCOUNT, SENDTYPE, RECVBUF, RECVCOUNT,
RECVTYPE, COMM, IERROR\fP)
<type> \fISENDBUF(*), RECVBUF(*)\fP
INTEGER \fISENDCOUNT, SENDTYPE, RECVCOUNT, RECVTYPE\fP
INTEGER \fICOMM, IERROR\fP
.fi
.SH INPUT PARAMETERS
.ft R
.TP 1.2i
sendbuf
Starting address of send buffer (choice).
.TP 1.2i
sendcount
Number of elements to send to each process (integer).
.TP 1.2i
sendtype
Datatype of send buffer elements (handle).
.TP 1.2i
recvcount
Number of elements to receive from each process (integer).
.TP 1.2i
recvtype
Datatype of receive buffer elements (handle).
.TP 1.2i
comm
Communicator over which data is to be exchanged (handle).
.SH OUTPUT PARAMETERS
.ft R
.TP 1.2i
recvbuf
Starting address of receive buffer (choice).
.ft R
.TP 1.2i
IERROR
Fortran only: Error status (integer).
.SH DESCRIPTION
.ft R
MPI_Neighbor_alltoall is a collective operation in which all processes send and receive the same amount of data to each neighbor. The operation of this routine can be represented as follows, where each process performs 2n (n being the number of neighbors in communicator \fIcomm\fP) independent point-to-point communications. The neighbors and buffer layout are determined by the topology of \fIcomm\fP.
.sp
Example of MPI_Neighbor_alltoall semantics for cartesian topologies:
.sp
.nf
MPI_Cart_get(\fIcomm\fP, maxdims, dims, periods, coords);
for (dim = 0, i = 0 ; dim < dims ; ++dim) {
MPI_Cart_shift(\fIcomm\fP, dim, 1, &r0, &r1);
MPI_Isend(\fIsendbuf\fP + i * \fIsendcount\fP * extent(\fIsendtype\fP),
\fIsendcount\fP, \fIsendtype\fP, r0, ..., \fIcomm\fP, ...);
MPI_Irecv(\fIrecvbuf\fP + i * \fIrecvcount\fP * extent(\fIrecvtype\fP),
\fIrecvcount\fP, \fIrecvtype\fP, r0, ..., \fIcomm\fP, ...);
++i;
MPI_Isend(\fIsendbuf\fP + i * \fIsendcount\fP * extent(\fIsendtype\fP),
\fIsendcount\fP, \fIsendtype\fP, r1, ..., \fIcomm\fP, &req[i]);
MPI_Irecv(\fIrecvbuf\fP + i * \fIrecvcount\fP * extent(\fIrecvtype\fP),
\fIrecvcount\fP, \fIrecvtype\fP, r1, ..., \fIcomm\fP, ...);
++i;
}
MPI_Waitall (...);
.fi
.sp
Each process breaks up its local \fIsendbuf\fP into n blocks \- each
containing \fIsendcount\fP elements of type \fIsendtype\fP \- and
divides its \fIrecvbuf\fP similarly according to \fIrecvcount\fP and
\fIrecvtype\fP. Process j sends the k-th block of its local
\fIsendbuf\fP to neighbor k, which places the data in the j-th block of
its local \fIrecvbuf\fP. The amount of data sent must be equal to the
amount of data received, pairwise, between every pair of processes.
.sp
.SH NEIGHBOR ORDERING
For a distributed graph topology, created with MPI_Dist_graph_create, the sequence of neighbors
in the send and receive buffers at each process is defined as the sequence returned by MPI_Dist_graph_neighbors
for destinations and sources, respectively. For a general graph topology, created with MPI_Graph_create, the order of
neighbors in the send and receive buffers is defined as the sequence of neighbors as returned by MPI_Graph_neighbors.
Note that general graph topologies should generally be replaced by the distributed graph topologies.
For a Cartesian topology, created with MPI_Cart_create, the sequence of neighbors in the send and receive
buffers at each process is defined by order of the dimensions, first the neighbor in the negative direction
and then in the positive direction with displacement 1. The numbers of sources and destinations in the
communication routines are 2*ndims with ndims defined in MPI_Cart_create. If a neighbor does not exist, i.e., at
the border of a Cartesian topology in the case of a non-periodic virtual grid dimension (i.e.,
periods[...]==false), then this neighbor is defined to be MPI_PROC_NULL.
If a neighbor in any of the functions is MPI_PROC_NULL, then the neighborhood collective communication behaves
like a point-to-point communication with MPI_PROC_NULL in this direction. That is, the buffer is still part of
the sequence of neighbors but it is neither communicated nor updated.
.sp
.SH NOTES
.ft R
The MPI_IN_PLACE option for \fIsendbuf\fP is not meaningful for this function.
.sp
All arguments on all processes are significant. The \fIcomm\fP argument,
in particular, must describe the same communicator on all processes. \fIcomm\fP
must be either a cartesian, graph, or dist graph communicator.
.sp
There are two MPI library functions that are more general than
MPI_Neighbor_alltoall. MPI_Neighbor_alltoallv allows all-to-all communication to and
from buffers that need not be contiguous; different processes may
send and receive different amounts of data. MPI_Neighbor_alltoallw expands
MPI_Neighbor_alltoallv's functionality to allow the exchange of data with
different datatypes.
.SH ERRORS
.ft R
Almost all MPI routines return an error value; C routines as
the value of the function and Fortran routines in the last argument.
.sp
Before the error value is returned, the current MPI error handler is
called. By default, this error handler aborts the MPI job, except for
I/O function errors. The error handler may be changed with
MPI_Comm_set_errhandler; the predefined error handler MPI_ERRORS_RETURN
may be used to cause error values to be returned. Note that MPI does not
guarantee that an MPI program can continue past an error.
.SH SEE ALSO
.ft R
.nf
MPI_Neighbor_alltoallv
MPI_Neighbor_alltoallw
MPI_Cart_create
MPI_Graph_create
MPI_Dist_graph_create
MPI_Dist_graph_create_adjacent