Estimating parallel speedup

Just Got Here

8:57:31 AM PST - Thu, Nov 22nd 2012
I've recently started running multi-host parallel NWCHEM runs. I would like to confirm what sort of scaling I am getting. Is it sufficient to look at the 'Task Times' as reported in the .out file & calculate the ratio of Cpu vs. Wall time as an estimate this? Or does one really have to run the same benchmark job on different number of nodes and compare the Wall time? I'm not looking for accurate figures - just a check to make sure I'm not squandering CPU!

Forum Vet

10:15:26 AM PST - Mon, Nov 26th 2012
The ratio of CPU vs Wall does not give you an idea of parallel performance. Differences between CPU and Wall give you insight into the effect of using disk, memory speed, etc. To assess parallel scalability you will have to run multiple jobs on different numbers of processors and compare Wall time. Bert Quote:Jarvist Nov 22nd 3:57 pm I've recently started running multi-host parallel NWCHEM runs. I would like to confirm what sort of scaling I am getting. Is it sufficient to look at the 'Task Times' as reported in the .out file & calculate the ratio of Cpu vs. Wall time as an estimate this? Or does one really have to run the same benchmark job on different number of nodes and compare the Wall time? I'm not looking for accurate figures - just a check to make sure I'm not squandering CPU!