The two additional functions introduced in this section can be used to implement all other topology functions. In general they will not be called by the user directly, unless he or she is creating additional virtual topology capability other than that provided by MPI. The two calls are both local.

MPI_CART_MAP(comm, ndims, dims, periods, newrank) | |

IN comm | input communicator (handle) |

IN ndims | number of dimensions of Cartesian structure (integer) |

IN dims | integer array of size ndims specifying the number of processes in each coordinate direction |

IN periods | logical array of size ndims specifying the periodicity specification in each coordinate direction |

OUT newrank | reordered rank of the calling process; MPI_UNDEFINED if calling process does not belong to grid (integer) |

` int MPI_Cart_map(MPI_Comm comm, int ndims, const int dims[], const int periods[], int *newrank) `

` MPI_Cart_map(comm, ndims, dims, periods, newrank, ierror) TYPE(MPI_Comm), INTENT(IN) :: comm INTEGER, INTENT(IN) :: ndims, dims(ndims) LOGICAL, INTENT(IN) :: periods(ndims) INTEGER, INTENT(OUT) :: newrank INTEGER, OPTIONAL, INTENT(OUT) :: ierror `

INTEGER COMM, NDIMS, DIMS(*), NEWRANK, IERROR

LOGICAL PERIODS(*)

MPI_CART_MAP computes an ``optimal'' placement for the calling process on the physical machine. A possible implementation of this function is to always return the rank of the calling process, that is, not to perform any reordering.

* Advice
to implementors.*

The function MPI_CART_CREATE(comm, ndims, dims, periods, reorder, comm_cart), with reorder = true can be implemented by calling MPI_CART_MAP(comm, ndims, dims, periods, newrank), then calling MPI_COMM_SPLIT(comm, color, key, comm_cart), with color = 0 if newrank MPI_UNDEFINED, color = MPI_UNDEFINED otherwise, and key = newrank. If ndims is zero then a zero-dimensional Cartesian topology is created.

The function MPI_CART_SUB(comm, remain_dims, comm_new) can be implemented by a call to MPI_COMM_SPLIT(comm, color, key, comm_new), using a single number encoding of the lost dimensions as color and a single number encoding of the preserved dimensions as key.

All other Cartesian topology functions can be implemented locally, using
the topology information that is cached with the communicator.
(* End of advice to implementors.*)

The corresponding function for
graph structures is as follows.

MPI_GRAPH_MAP(comm, nnodes, index, edges, newrank) | |

IN comm | input communicator (handle) |

IN nnodes | number of graph nodes (integer) |

IN index | integer array specifying the graph structure, see MPI_GRAPH_CREATE |

IN edges | integer array specifying the graph structure |

OUT newrank | reordered rank of the calling process; MPI_UNDEFINED if the calling process does not belong to graph (integer) |

` int MPI_Graph_map(MPI_Comm comm, int nnodes, const int index[], const int edges[], int *newrank) `

` MPI_Graph_map(comm, nnodes, index, edges, newrank, ierror) TYPE(MPI_Comm), INTENT(IN) :: comm INTEGER, INTENT(IN) :: nnodes, index(nnodes), edges(*) INTEGER, INTENT(OUT) :: newrank INTEGER, OPTIONAL, INTENT(OUT) :: ierror `

INTEGER COMM, NNODES, INDEX(*), EDGES(*), NEWRANK, IERROR

* Advice
to implementors.*

The function MPI_GRAPH_CREATE(comm, nnodes, index, edges, reorder, comm_graph), with reorder = true can be implemented by calling MPI_GRAPH_MAP(comm, nnodes, index, edges, newrank), then calling MPI_COMM_SPLIT(comm, color, key, comm_graph), with color = 0 if newrank MPI_UNDEFINED, color = MPI_UNDEFINED otherwise, and key = newrank.

All other graph topology functions can be implemented locally, using the
topology information that is cached with the communicator.
(* End of advice to implementors.*)

Return to MPI-3.1 Standard Index

Return to MPI Forum Home Page

(Unofficial) MPI-3.1 of June 4, 2015

HTML Generated on June 4, 2015