Older results:
Nodes 1 2 4 8 16 32 64 128 256 512
IBM/SP2 1.93 0.98 0.50 0.26 0.14 0.09
The results in summary page are produced
on an IBM/SP2 machine
These are some results with the older version of CHARMM
Node Eext Eint Wait Comm List Integ Total Speedup eff.
1 5654.3 60.5 0.0 0.0 1227.4 9.4 6951.7=1.93h 1.0 100%
2 2815.6 33.1 22.2 21.5 615.9 5.0 3513.4=0.98h 1.98 99%
4 1416.0 19.5 13.5 33.5 310.3 2.6 1795.4=0.50h 3.87 97%
8 720.1 12.6 10.2 40.8 157.0 1.5 942.2=0.26h 7.38 92%
16 372.4 9.1 7.9 45.6 80.4 1.2 516.5=0.14h 13.46 84%
32 197.8 7.1 6.0 67.3 42.0 1.3 321.4=0.09h 21.63 68%
CPU-time (efficiency %)
Nodes: 1 2 4 8 16 32
Benchmark -----------------------------------------------
NIH Mb+water 3.8 1.9/100 0.91/100 0.51/92 0.28/84 0.14/84 (hours/%)
CMU Mb+water 18 9.7/93 5.9/76 4.2/54 2.7/42 2.0/28 (minutes/%)
the CMU benchmark uses shorter cutoffs and has ca 175 nbonds/atom
whereas NIH has 477 nbonds/atom.
See IBM WWW pages for more details
From: "Lennart Nilsson, KI/CSB" (ln@csb.ki.se)
if anybody is interested I have scripts etc to run charmm under POE and the LoadLeveler.