I am struggling with a build of NWChem on an Xeon Cluster with EDR IB getting very different performance (4x) on a medium CCSD model when using different MPIs (Intel MPI 17 and HPE MPT). The only thing different in the output files other than timing numbers is the following lines:
< Intel MPI > HPE MPT
264 < !! scf_movecs_sym_adapt: 40 vectors were symmetry contaminated
265 ---
266 > !! scf_movecs_sym_adapt: 38 vectors were symmetry contaminated
.........
280 < Record size in doubles = 65536 No. of integs per rec = 43688
281 < Max. records in memory = 2 Max. records in file = 140004
282 < No. of bits per label = 8 No. of bits per value = 64
283 ---
284 > Record size in doubles = 65536 No. of integs per rec = 43688
285 > Max. records in memory = 2 Max. records in file = 6433769
286 > No. of bits per label = 8 No. of bits per value = 64
Can anyone give me a clue what causes the "Max. records in file" to be so different in the 2 cases and if that could be contributing to a serious performance difference between the 2 runs?
Thanks
Jan Fredin
|