I attempt to benchmark the performance difference of CCSD calculations on KNL system with different MCDRAM modes:
I compiled 2 copies of NWChem 6.8 on our KNL system; the compilation environment are identical except that one copy has USE_FASTMEM=T and the other one export USE_FASTMEM=F:
I managed to complete a (H2O)4 CCSD(T)/cc-pvqz (TCE) calculation for the copy using USE_FASTMEM=T on the nodes with MCDRAM in flat mode, while another copy ( USE_FASTMEM=F) fail on the nodes which have MCDRAM in cache mode.
Can anyone advice how to fix this problem? Thanks!
number of tasks: 11984
1 ga offset 1570646329 size_xx_perproc 1570646329mx 4
2 ga offset 3141292658 size_xx_perproc 1570646329mx 4
3 ga offset 4711938987 size_xx_perproc 1570646331mx 4
Cpu & wall time / sec 481.0 516.7
T1-number-of-tasks 13
t1 file size = 7040
t1 file name = ./rubbish.w4.t1
t1 file handle = -998
T2-number-of-boxes 260
get-block-ind1 2
------------------------------------------------------------------------
get-block-ind1 2
------------------------------------------------------------------------
------------------------------------------------------------------------
get-block-ind1 2
------------------------------------------------------------------------
------------------------------------------------------------------------
------------------------------------------------------------------------
get-block-ind1 2
------------------------------------------------------------------------
------------------------------------------------------------------------
current input line :
current input line :
current input line :
|