Enrich starpu_mpi_data_cpy by specifying the destination
Is your feature request related to a problem? Please describe.
We would like a simpler way to trigger a copy of a data from one node to a given list of other nodes (without creating a new communicator every time).
Describe the solution you'd like
Send from processor 0 to 1 and 3 :
Register a data "workspace_buffer" that is present on all processes (using per_node flag).
then call starpu_mpi_data_cpy_dest(src_data_on_proc0, workspace_buffer, comm, async, callback_func, callback_arg, dest );
Describe alternatives you've considered Have a different handle for each handle that will receive the data. It can be very verbose but should work.