mhp MPI backend

[lslusarczyk]

Summary

MPI based multi process communication primitives as a backend for Distributed-Ranges.

Feature details

MPI backend allows to write programs in SPMD way. Data structures are automatically distributed on multiple nodes using MPI and their constructors and algorithms are collective.
The MPI backedn will use a relaxed memory model similar to OpenSHMEM in which writes return without blocking while reads block until the data is ready to be read. To ensure consistency, collective operations perform a barrier/memory flush to ensure
all writes are complete before returning control. Collective operations employ asynchronous communication internally.

The feature will provide an MPI backend (communication primitives implemented using MPI) which can be used as a backend for

• distributed_vector
• distributed_mdarray
• algorithms: for_each, for_each_tile, copy, fill, reduce, sort, inclusive_scan, exclusive_scan, transform,
• Distributed Range Views: zip, iota, take, drop, counted, enumerate, transform
• Distributed Mdspan Algorithms: mdfor_each, mdfor_each_stencil, mdfor_each_tile

MPI Communicatin primitives should support:

• copy to/from local and Distributed Range
• implementing a distributed range remote reference, which is exposed as [index]
• implementing distributed mdspan remote reference, which is exposed as (index0, index1, ...)
• halo exchange in distributed_vector and distributed_mdarray