MPI barriers in IPE?

[rmontuoro]

A few MPI barriers are used in the refactored IPE code.

We should discuss whether such barriers are required or could be eliminated since they affect runtime performance.

[gmillward]

Is this a problem with IPE running coupled to WAM? If it's just a performance issue then this would seem low-priority, unless the performance is terrible for some reason?

[rmontuoro]

It is a low-priority issue, indeed. However, it should be addressed before code delivery since MPI barriers do affect runtime performance and may mask code issues when run in parallel.

[gmillward]

Rafaelle - looking at the code, MPI_barrier is only used (during actual runtime) for writing out the IPE hdf5 'state' files. I think the barrier is in there because Joe had problems with the hdf5 files (ie, they were bogus/corrupted) when there was no barrier. I seem to recall Joe found a few bugs in hdf5 itself....but yeh, if it's a real performance issue then we can see what can be done.

[twfang]

Is it solved? should we close it?

[rmontuoro]

Active MPI barriers have been eliminated after introducing the updated HDF5 I/O layer in IPE.

One barrier still remain in the unused legacy subroutine Write_Electron_Density_to_HDF5 in IPE_Plasma_Class.F90.
Can we delete this legacy subroutine, or should it be updated to use the HDF5 I/O later?

[rmontuoro]

Write_Electron_Density_to_HDF5() is not used in IPE and should be removed.