Some performance measurements. I added on a temporary branch some ENV variables to enable/disable the included changes on this PR, and these are the results for a simple 15GB compartment report and 80 processes (2 nodes):
GPFS (no changes): 8.88757s
IME (FUSE no changes): 5.274s
IME (FUSE + MPI Hints): 3.75662s
IME (IME Native + MPI Hints): 1.60297s
Enabling the IME native API without the hints does not seem to help compared to using FUSE, as suggested by DDN's team. Also, disabling the collective buffering and data sieving (i.e., MPI Hints in the results) does not help GPFS and the execution time degrades to 38.8573s, as expected.
Some performance measurements. I added on a temporary branch some ENV variables to enable/disable the included changes on this PR, and these are the results for a simple 15GB compartment report and 80 processes (2 nodes):
Enabling the IME native API without the hints does not seem to help compared to using FUSE, as suggested by DDN's team. Also, disabling the collective buffering and data sieving (i.e., MPI Hints in the results) does not help GPFS and the execution time degrades to 38.8573s, as expected.