On 23 May 2011 21:15, Martinelli Fabio <[email protected]> wrote: > Dear Mark > > I tried that parameter in sge6.2u5 but it doesn't work and the cpu seems to > be = ru_utime + ru_stime Can I ask how you determined that it doesn't work? It looks like you are assuming that the sharetree and qacct -j always use the same definition of resource usage. I think a fair experiment would look something like this: On an empty cluster with all projects allocated the same fairshare and no significant history. user/project 1 submits long sleeper jobs user/project 2 submits an equal number of short cpu intensive jobs.
See who gets the most fair share tickets. > > instead by replacing SHARETREE_RESERVED_USAGE=true with > ACCT_RESERVED_USAGE=true > > I got a 10s sleep job ( doing anything ) correctly accounted for 10s like > showed in this qacct output: > $ qacct -j 1151|egrep 'ru_wallclock|cpu' > ru_wallclock 10 > cpu 10.000 I don't think the docs claim that the output from qacct -j will always be the usage as per the share tree. The problem with enabling ACCT_RESERVED_USAGE is that it means you can't use Grid Engine logs to determine if your jobs are running efficiently. > > now I hope sge security mechanisms prevent jobs to consume more CPU time than > their Wallclock time, I'm studying this topic too.. > > thanks again for your support > regards > Fabio > > _______________________________________________ > users mailing list > [email protected] > https://gridengine.org/mailman/listinfo/users > _______________________________________________ users mailing list [email protected] https://gridengine.org/mailman/listinfo/users
