Implement standard MPI ABI #43
Conversation
|
@eschnett I'm a bit curious as to what the status of this is? Is it a PoC or will it make it to a release? |
|
This is beta quality, so to say. It works locally with MPICH and OpenMPI on Linux and macOS. I'll have to check with the current draft of the next MPI standard (to avoid accidental incompatibilities), clean it up a bit, and then start testing on actual HPC systems. It's delayed because I am busy with other projects, but this is definitively the future of MPItrampoline. |
|
... and test with a few real-world applications, of course. |
|
@eschnett Now that the ABI standard is in place, do you think you will continue this? I ask because I very much like the mechanism that MPItrampoline has for injecting a different implementation. We use rpath a lot these days and I feel like the environment variable approach used here is very well suited to a kind of "MPI-settings" environment module where you use environment variables to select and tune the MPI backend (without affecting the software stack in any way). |
|
@ocaisa Yes, I definitely plan on continuing this. At the moment I have investigated building MPICH with the new ABI (this appears to be supported already but I need to check how closely it conforms to the standard). The question remains what to do with Fortran. The MPI ABI does not specify an ABI for Fortran. Instead, the idea is to write a Fortran wrapper for the new ABI that implements the Fortran MPI standard. This could be a stand-alone library. |
2 participants
No description provided.