Topology Constructors

Cartcomm Intracomm.Create_cart(int [] dims, boolean [] periods,
                               boolean reorder)

dims	the number of processes in each dimension
periods	true if grid is periodic, false if not, in each dimension
reorder	true if ranking may be reordered, false if not
returns:	new Cartesian topology communicator

Create a Cartesian topology communicator whose group is a subset of the group of this communicator. Java binding of the MPI operation MPI_CART_CREATE. The number of dimensions of the Cartesian grid is taken to be the size of the dims argument. The array periods must be the same size.

static Cartcomm.Dims_create(int nnodes, int [] dims)

nnodes	number of nodes in a grid
dims	array specifying the number of nodes in each dimension

Select a balanced distribution of processes per coordinate direction. Java binding of the MPI operation MPI_DIMS_CREATE. Number of dimensions is the size of is dims. Note that dims is an inout parameter.

Graphcomm Intracomm.Create_graph(int [] index, int [] edges,
                                 boolean reorder)

index	node degrees
edges	graph edges
reorder	true if ranking may be reordered, false if not
returns:	new graph topology communicator

Create a graph topology communicator whose group is a subset of the group of this communicator. Java binding of the MPI operation MPI_GRAPH_CREATE. The number of nodes in the graph, nnodes, is taken to be size of the index argument. The size of array edges must be $\mbox{\tt index}[\mbox{\em nnodes} - 1]$.

int Comm.Topo_test()

returns:

topology type of communicator

Returns the type of topology associated with the communicator. Java binding of the MPI operation MPI_TOPO_TEST. The return value will be one of MPI.GRAPH, MPI.CART or MPI.UNDEFINED.

GraphParms Graphcomm.Get()

returns:

object defining node degress and edges of graph

Returns graph topology information. Java binding of the MPI operations MPI_GRAPHDIMS_GET and MPI_GRAPH_GET. The class of the returned object is

  public class GraphParms {
    public int [] index ;
    public int [] edges ; 
  }

The number of nodes and number of edges can be extracted from the sizes of the index and edges arrays.

CartParms Cartcomm.Get()

returns:

object containing dimensions, periods and local coordinates

Returns Cartesian topology information. Java binding of the MPI operations MPI_CARTDIM_GET and MPI_CART_GET. The class of the returned object is

  public class CartParms {
    public int [] dims ;
    public booleans [] periods ;
    public int [] coords ; 
  }

The number of dimensions can be obtained from the size of (eg) the dims array.

• Rationale. The inquiries MPI_GRAPHDIMS_GET, MPI_GRAPH_GET, MPI_CARTDIM_GET, and MPI_CART_GET are unusual in returning multiple independent values from single calls. This is a problem in Java. The Java binding could split these inquiries into several independent ones, but that would complicate JNI-based wrapper implementations. Hence we introduced the auxilliary classes GraphParms and CartParms to hold multiple results.(End of rationale.)

int Cartcomm.Rank(int [] coords)

coords	Cartesian coordinates of a process
returns:	rank of the specified process

Translate logical process coordinates to process rank. Java binding of the MPI operation MPI_CART_RANK.

int [] Cartcomm.Coords(int rank)

coords	rank of a process
returns:	Cartesian coordinates of the specified process

Translate process rank to logical process coordinates. Java binding of the MPI operation MPI_CART_COORDS.

int [] Graphcomm.Neighbours(int rank)

coords	rank of a process in the group of this communicator
returns:	array of ranks of neighbouring processes to one specified

Provides adjacency information for general graph topology. Java binding of the MPI operations MPI_GRAPH_NEIGHBOURS_COUNT and MPI_GRAPH_NEIGHBOURS. The number of neighbours can be extracted from the size of the result.

ShiftParms Cartcomm.Shift(int direction, int disp)

direction	coordinate dimension of shift
disp	displacement
returns:	object containing ranks of source and destination processes

Compute source and destination ranks for ``shift'' communication. Java binding of the MPI operation MPI_CART_SHIFT. The class of the returned object is

  public class ShiftParms {
    public int rankSource ;
    public int rankDest ; 
  }

Cartcomm Cartcomm.Sub(boolean [] remainDims)

remainDims	by dimension, true if dimension is to be kept, false otherwise
returns:	communicator containing subgrid including this process

Partition Cartesian communicator into subgroups of lower dimension. Java binding of the MPI operation MPI_CART_SUB.

int Cartcomm.Map(int [] dims, boolean [] periods)

dims	the number of processes in each dimension
periods	true if grid is periodic, false if not, in each dimension
returns:	reordered rank of calling process

Compute an optimal placement. Java binding of the MPI operation MPI_CART_MAP. The number of dimensions is taken to be size of the dims argument.

int Graphcomm.Map(int [] index, int [] edges)

index	node degrees
edges	graph edges
returns:	reordered rank of calling process

Compute an optimal placement. Java binding of the MPI operation MPI_GRAPH_MAP. The number of nodes is taken to be size of the index argument.