Topology

Cartesian

MPI.Dims_createFunction
newdims = Dims_create(nnodes::Integer, dims)

A convenience function for selecting a balanced Cartesian grid of a total of nnodes nodes, for example to use with MPI.Cart_create.

dims is an array or tuple of integers specifying the number of nodes in each dimension. The function returns an array newdims of the same length, such that if newdims[i] = dims[i] if dims[i] is non-zero, and prod(newdims) == nnodes, and values newdims are as close to each other as possible.

nnodes should be divisible by the product of the non-zero entries of dims.

External links

source
MPI.Cart_createFunction
comm_cart = Cart_create(comm::Comm, dims; periodic=map(_->false, dims), reorder=false)

Create new MPI communicator with Cartesian topology information attached.

dims is an array or tuple of integers specifying the number of MPI processes in each coordinate direction, and periodic is an array or tuple of Bools indicating the periodicity of each coordinate. prod(dims) must be less than or equal to the size of comm; if it is smaller than some processes are returned a null communicator.

If reorder == false then the rank of each process in the new group is identical to its rank in the old group, otherwise the function may reorder the processes.

See also MPI.Dims_create.

External links

source
MPI.Cart_getFunction
dims, periods, coords = Cart_get(comm::Comm)

Obtain information on the Cartesian topology of dimension N underlying the communicator comm. This is specified by two Cint arrays of N elements for the number of processes and periodicity properties along each Cartesian dimension. A third Cint array is returned, containing the Cartesian coordinates of the calling process.

External links

source
MPI.Cart_coordsFunction
coords = Cart_coords(comm::Comm, rank::Integer=Comm_rank(comm))

Determine coordinates of a process with rank rank in the Cartesian communicator comm. If no rank is provided, it returns the coordinates of the current process.

Returns an integer array of the 0-based coordinates. The inverse of Cart_rank.

External links

source
MPI.Cart_rankFunction
rank = Cart_rank(comm::Comm, coords)

Determine process rank in communicator comm with Cartesian structure. The coords array specifies the 0-based Cartesian coordinates of the process. This is the inverse of MPI.Cart_coords

External links

source
MPI.Cart_shiftFunction
rank_source, rank_dest = Cart_shift(comm::Comm, direction::Integer, disp::Integer)

Return the source and destination ranks associated to a shift along a given direction.

External links

source
MPI.Cart_subFunction
comm_sub = Cart_sub(comm::Comm, remain_dims)

Create lower-dimensional Cartesian communicator from existent Cartesian topology.

remain_dims should be a boolean vector specifying the dimensions that should be kept in the generated subgrid.

External links

source
MPI.Cartdim_getFunction
ndims = Cartdim_get(comm::Comm)

Return number of dimensions of the Cartesian topology associated with the communicator comm.

External links

source

Graph topology

MPI.Dist_graph_createFunction
graph_comm = Dist_graph_create(comm::Comm, sources::Vector{Cint}, degrees::Vector{Cint}, destinations::Vector{Cint}; weights::Union{Vector{Cint}, Unweighted, WeightsEmpty}=UNWEIGHTED, reorder=false, infokws...)

Create a new communicator from a given directed graph topology, described by incoming and outgoing edges on an existing communicator.

Arguments

  • comm::Comm: The communicator on which the distributed graph topology should be induced.
  • sources::Vector{Cint}: An array with the ranks for which this call will specify outgoing edges.
  • degrees::Vector{Cint}: An array with the number of outgoing edges for each entry in the sources array.
  • destinations::Vector{Cint}: An array containing destination nodes for the source nodes in the source node array, of lengthsum(sources).
  • weights::Union{Vector{Cint}, Unweighted, WeightsEmpty}: The edge weights of the specified edges. The default is MPI.UNWEIGHTED.
  • reorder::Bool=false: If set true, then the MPI implementation can reorder the source and destination indices.

Example

We can generate a ring graph 1 --> 2 --> ... --> N --> 1, where N is the number of ranks in the communicator, as follows

julia> rank = MPI.Comm_rank(comm);
julia> N = MPI.Comm_size(comm);
julia> sources = Cint[rank];
julia> degrees = Cint[1];
julia> destinations = Cint[mod(rank-1, N)];
julia> graph_comm = Dist_graph_create(comm, sources, degrees, destinations)

External links

source
MPI.Dist_graph_create_adjacentFunction
graph_comm = Dist_graph_create_adjacent(comm::Comm,
    sources::Vector{Cint}, destinations::Vector{Cint};
    source_weights::Union{Vector{Cint}, Unweighted, WeightsEmpty}=UNWEIGHTED, destination_weights::Union{Vector{Cint}, Unweighted, WeightsEmpty}=UNWEIGHTED,
    reorder=false, infokws...)

Create a new communicator from a given directed graph topology, described by local incoming and outgoing edges on an existing communicator.

Arguments

  • comm::Comm: The communicator on which the distributed graph topology should be induced.
  • sources::Vector{Cint}: The local, incoming edges on the rank of the calling process.
  • destinations::Vector{Cint}: The local, outgoing edges on the rank of the calling process.
  • source_weights::Union{Vector{Cint}, Unweighted, WeightsEmpty}: The edge weights of the local, incoming edges. The default is MPI.UNWEIGHTED.
  • destinations_weights::Union{Vector{Cint}, Unweighted, WeightsEmpty}: The edge weights of the local, outgoing edges. The default is MPI.UNWEIGHTED.
  • reorder::Bool=false: If set true, then the MPI implementation can reorder the source and destination indices.

Example

We can generate a ring graph 1 --> 2 --> ... --> N --> 1, where N is the number of ranks in the communicator, as follows

julia> rank = MPI.Comm_rank(comm);
julia> N = MPI.Comm_size(comm);
julia> sources = Cint[mod(rank-1, N)];
julia> destinations = Cint[mod(rank+1, N)];
julia> graph_comm = Dist_graph_create_adjacent(comm, sources, destinations);

External links

source
MPI.Dist_graph_neighbors_countFunction
indegree, outdegree, weighted = Dist_graph_neighbors_count(graph_comm::Comm)

Return the number of in and out edges for the calling processes in a distributed graph topology and a flag indicating whether the distributed graph is weighted.

Arguments

  • graph_comm::Comm: The communicator of the distributed graph topology.

Example

Let us assume the following graph 0 <--> 1 --> 2, which has no weights on its edges, then the process with rank 1 will obtain the following result from calling the function

julia> Dist_graph_neighbors_count(graph_comm)
(1,2,false)

External links

source
MPI.Dist_graph_neighbors!Function
Dist_graph_neighbors!(graph_comm::MPI.Comm,
   sources::Vector{Cint}, source_weights::Union{Vector{Cint}, Unweighted},
   destinations::Vector{Cint}, destination_weights::Union{Vector{Cint}, Unweighted},
)
Dist_graph_neighbors!(graph_comm::Comm, sources::Vector{Cint}, destinations::Vector{Cint})

Query the neighbors and edge weights (optional) of the calling process in a distributed graph topology.

Arguments

  • graph_comm::Comm: The communicator of the distributed graph topology.
  • sources: A preallocated Vector{Cint}, which will be filled with the ranks of the processes whose edges pointing towards the calling process. The length is exactly the indegree returned by MPI.Dist_graph_neighbors_count.
  • source_weights: A preallocated Vector{Cint}, which will be filled with the weights associated to the edges pointing towards the calling process. The length is exactly the indegree returned by MPI.Dist_graph_neighbors_count. Alternatively, MPI.UNWEIGHTED can be used if weight information is not required.
  • destinations: A preallocated Vector{Cint}, which will be filled with the ranks of the processes towards which the edges of the calling process point. The length is exactly the outdegree returned by [MPI.Distgraphneighbors_count`](@ref).
  • destination_weights: A preallocated Vector{Cint}, which will be filled with the weights associated to the edges of the outgoing edges of the calling process point. The length is exactly the outdegree returned by MPI.Dist_graph_neighbors_count. Alternatively, MPI.UNWEIGHTED can be used if weight information is not required.

Example

Let us assume the following graph:

            rank 0 <-----> rank 1 ------> rank 2
weights:              3             4

then then the process with rank 1 will need to preallocate sources and source_weights as vectors of length 1, and a destinations and destination_weights as vectors of length 2.

The call will fill the vectors as follows:

julia> MPI.Dist_graph_neighbors!(graph_comm, sources, source_weights, destinations, destination_weights);
julia> sources
[0]
julia> source_weights
[3]
julia> destinations
[0,2]
julia> destination_weights
[3,4]

Note that the edge between ranks 0 and 1 can have a different weight depending on whether it is the incoming edge 0 --> 1 or the outgoing one 0 <-- 1.

See also

External links

source
MPI.Dist_graph_neighborsFunction
sources, source_weights, destinations, destination_weights = Dist_graph_neighbors(graph_comm::MPI.Comm)

Return (sources, source_weights, destinations, destination_weights) of the graph communicator graph_comm. For unweighted graphs source_weights and destination_weights are returned as MPI.UNWEIGHTED.

This function is a wrapper around MPI.Dist_graph_neighbors_count and MPI.Dist_graph_neighbors! that automatically handles the allocation of the result vectors.

source