UMUI panel subindep_SubmitMethod and subindep_ResQsub differs XC30 and XC40

[markr]

Hello CMS,
using the XC40 (Monsoon) I see that the qsub panel offers different choices to that of the XC30 (Archer) I presume that is due to the differing number of cores per node (24 archer and 32 monsoon).

I would like to be able to control this value for Monsoon as well so that I can show a "pure" OpenMP effect for 1,2,3,4,5,6,7,8 threads while keeping the MPI distribution the same.

Also I note that the advice is to activate hyperthreads. I have found this detrimental to the cases I use by about 20% slower.
This also interferes with the OpenMP performance.
When I deactivate HT then the code takes less time and the OpenMP performance is closer to ideal for the sections I am analyzing.

Have you any advice on how I could provide a variation on the aprun command? I presume it is set per version per machine within the um account scripts.

many thanks,
Mark

[ros]

Hi Mark,

The options differ between the XC40 & XC30 primarily because the Met Office configure their options in a different way to how we did on ARCHER. Since MONSooN is supposed to mirror the Met Office the options for the XC40 are the same as how the Met office run their vn8.x jobs internally.

You can still control the number of cores per node for XC40 but you will have to modify the aprun command in the qsatmos script either via a branch or you could hack the $DATAW/bin/qsatmos on MONSOON, but be aware if you hack it on MONSooN it will get overwritten if you submit to compile.

Cheers,
Ros.

[markr]

Thank you Ros.
I will try to modify the behaviour with the advice I got for Cray PAT compilation.