A (vertex-centric) distributed communication graph. Each vertex of the graph corresponds to a rank and each edge (i,j) connect two ranks i and j which can communicate with each other. The distributed communication graph is vertex-centric in the sense that on each rank the local graph only contains the corresponding vertex and its in and out neighborhood. Note that MPI allow multiple edges between the same ranks i and j, i.e. the distributed communication graph can be a multi-graph.
More...
#include <distributed_graph_communicator.hpp>
|
|
| DistributedCommunicationGraph ()=default |
| | Default constructor.
|
| |
| template<typename InNeighborsRange , typename OutNeighborsRange > |
| | DistributedCommunicationGraph (InNeighborsRange const &in_neighbors, OutNeighborsRange const &out_neighbors) |
| | Constructs a communication graph based on a range of in-going and out-going edges which might be weighted. A unweighted edge is simply an integer, wherea a weighted edge is a pair-like object consisting of two integers which can be decomposed by a structured binding via auto [rank, weight] = weighted_edge;.
|
| |
|
| DistributedCommunicationGraph (DefaultContainer< int > &&in_ranks, DefaultContainer< int > &&out_ranks) |
| | Constructs a communication graph based on a range of unweighted in-going and out-going neighbors, i.e., ranks.
|
| |
|
| DistributedCommunicationGraph (DefaultContainer< int > &&in_ranks, DefaultContainer< int > &&out_ranks, DefaultContainer< int > &&in_weights, DefaultContainer< int > &&out_weights) |
| | Constructs a communication graph based on a range of weighted in-going and out-going neighbors. The ownership of the underlying ranks/weights containers is transfered.
|
| |
|
template<typename NeighborRange > |
| | DistributedCommunicationGraph (NeighborRange const &neighbors) |
| | Constructs a communication graph based where in and out neighbors are the same, i.e. a symmetric neighborhood/graph.
|
| |
|
CommunicationGraphLocalView | get_view () const |
| | Returns a view of the communication graph.
|
| |
| template<typename Map = std::unordered_map<size_t, size_t>> |
| auto | get_rank_to_out_neighbor_idx_mapping () |
| | In neighborhood collectives the order of sent and received data depends on the ordering of the underlying out and in neighbors (note that the MPI standard allows that a neighbor occurs multiple times within the neighbors list). For example in MPI_Neighbor_alltoall the k-th block in the send buffer is sent to the k-th neighboring process. Hence, to exchange data to rank r via neighborhood collectives it might be useful to know the index of rank r within the out neighbors (provided r is a neighbor at all). Therefore, this function returns a mapping from the rank of each out neighbor r to its index within the out neighbors. If r occurs multiple times, one of these positions is returned in the mapping.
|
| |
template<
template< typename... >
typename DefaultContainer = std::vector>
class kamping::DistributedCommunicationGraph< DefaultContainer >
A (vertex-centric) distributed communication graph. Each vertex of the graph corresponds to a rank and each edge (i,j) connect two ranks i and j which can communicate with each other. The distributed communication graph is vertex-centric in the sense that on each rank the local graph only contains the corresponding vertex and its in and out neighborhood. Note that MPI allow multiple edges between the same ranks i and j, i.e. the distributed communication graph can be a multi-graph.
◆ DistributedCommunicationGraph()
Constructs a communication graph based on a range of in-going and out-going edges which might be weighted. A unweighted edge is simply an integer, wherea a weighted edge is a pair-like object consisting of two integers which can be decomposed by a structured binding via auto [rank, weight] = weighted_edge;.
- Template Parameters
-
| InNeighborsRange | Range type of in-going neighbors. |
| OutNeighborsRange | Range type of out-going neighbors. |
- Parameters
-
| in_neighbors | Range of in-going neighbors. |
| out_neighbors | Range of out-going neighbors. |
◆ get_rank_to_out_neighbor_idx_mapping()
template<
typename Map = std::unordered_map<size_t, size_t>>
In neighborhood collectives the order of sent and received data depends on the ordering of the underlying out and in neighbors (note that the MPI standard allows that a neighbor occurs multiple times within the neighbors list). For example in MPI_Neighbor_alltoall the k-th block in the send buffer is sent to the k-th neighboring process. Hence, to exchange data to rank r via neighborhood collectives it might be useful to know the index of rank r within the out neighbors (provided r is a neighbor at all). Therefore, this function returns a mapping from the rank of each out neighbor r to its index within the out neighbors. If r occurs multiple times, one of these positions is returned in the mapping.
- Template Parameters
-
| Map | Map type storing the mapping, uses std::unordered_map<size_t, size_t> as default. |
- Returns
- Returns the mapping.
The documentation for this class was generated from the following file:
1.10.0