Speed up engine_exchange_cells

Changes:

• Split the different exchanges into separate loops over proxies.
• Made all MPI calls on thread 0.
• Add threadpool loops trying to progress the comms.
• Parallelized the cell unpacking.

added enhancement label

assigned to @matthieu

added 1 commit

• fedb340f - Make the first exchange use a single thread

Compare with previous version

added 1 commit

• 6e491cfd - Split all the calls in the cell exchange

Compare with previous version

added 1 commit

• f1e9e80b - Code formatting

Compare with previous version

added 1 commit

• ef626ab6 - Use a threadpool for the cell unpacking. Use a threadpool to help MPI progress...

Compare with previous version

added 1 commit

• 4c13f79f - Code formatting

Compare with previous version

added MPI label

added 6 commits

• 4c13f79f...cacd3a1e - 5 commits from branch master
• 7ed5db9e - Merge branch 'master' into cell_exchange_improvements

Compare with previous version

changed the description

added 2 commits

• dc6ee6a1 - 1 commit from branch master
• 1c7de991 - Merge branch 'master' into cell_exchange_improvements

Compare with previous version

added 4 commits

• 1c7de991...22adecfe - 2 commits from branch master
• 040bab5f - Merge branch 'master' into cell_exchange_improvements
• fd3fc134 - Remove unwanted suppression of check

Compare with previous version

added 2 commits

• dc75df4d - Better documentation of the logic of the exchanges
• 5a1e537f - Activate the verbose documenting of all the particle exchanges only when a specific define is set.

Compare with previous version

added 1 commit

• 97d6a7a1 - Activate the verbose documenting of all the particle exchanges only when a specific define is set.

Compare with previous version

added 1 commit

• f2d19659 - Better naming and documentation of the operations in exchange_cells

Compare with previous version

added 1 commit

• 3b4dd24f - Parallelize the proxy's memcpy over the threadpool

Compare with previous version

added 1 commit

• fa5b4900 - Rename the operations in the cell_exchange to better reflect what the various mappers actually are

Compare with previous version

added 1 commit

• 94f79d4a - Emit the receives before the sends

Compare with previous version

Open questions:

• Would it help to keep the proxy memory allocated permanently?
• Do we need to transfer the multipoles already at this point?

added 40 commits

• 94f79d4a...e01bea72 - 39 commits from branch master
• 14ccb086 - Merge branch 'master' into cell_exchange_improvements

Compare with previous version