Resurrect the MPI version of VAMP

[kilian]

Regarding VAMP, this is probably fully functional already, but some work may have to be done on the WHIZARD side.

This would nicely complement the OpenMP implementation in the innermost loop (the sampling function).

[Juergen Reuter]

BTW, is there any update on this?

[kilian]

Update from MT: the MPI version is complete and compiles, but VAMP(I) doesn't produce correct numbers yet. Debugging.

[Juergen Reuter]

After reports from Siegen, this might hopefully be ready earlier....

[Juergen Reuter]

Guess this job moved now to SB.

[Juergen Reuter]

Milestone renamed

[kilian]

The actual re-implementation has been done by SB.

Now, the code has to be merged with the master branch, but there are some open questions about the top-level management of the MPI workers in the context of the WHIZARD workflow. Profiling would also be useful.

There is significant and linear speedup, even for 100 cores, provided the number of calls for channel is sufficiently large.

For parameter sets with a small number of calls per channel, there is saturation. A possible solution is indicated in ticket #818

[Juergen Reuter]

Milestone renamed

[Juergen Reuter]

This has been merged (fully) with the master. Documentation is under way, but still pending and not yet in the master. Benchmarking is also in the making. Shall we close the ticket, as there are already more specialised tickets like #818?

[Juergen Reuter]

After discussion with SB, this ticket will be happily closed. All technical improvements and benchmarking as well as documentation will be part of specialized tickets.