Closed ewels closed 9 years ago
@s-andrews: I assume GRIDEngine has some kind of similar support? Would you be interested in adding this for SGE?
We have some nice resource usage tools which I'm might try to use to optimise module requests. I'll try to write some sensible time estimation code for modules that I use as well, rather than just flat estimates. Obviously setting these too low will be very annoying if jobs get cancelled, the increased queue efficiency should be worth it though.
GridEngine has qacct
which records some stats for each job run (CPU, runtime and memory mostly). We've been tweaking the existing limits to match what we see from real jobs as a lot of the older limits proved to be a little under where they needed to be.
I'm sure GridEngine can be set up to take runtime into account when submitting jobs (normally I think people just set up short and long queues with different limits on them), but it's not something we use.
Ok, sounds good - if there are any new limit tweaks you've done recently, it'd be good to see them to give me a head start on fine tuning all of this. I'll implement the time estimation thing as a parameter for the qsub
submission string, so it'll be totally optional and there if you want it in the future.
OK, will check if we have any unsubmitted tweaks.
I was having problem getting our branch back in line with your mainline version. I couldn't generate a pull which seemed to bring them back together. If you had a second to see if you could figure it out that would help us out (I think you have commit privileges on our branch).
Simon.
The PR looks like there's only one change you've done from the 0.4 devel branch, adding requirements to cf_download
. There's a merge conflict though, so I've probably done something similar at my end.
Not sure it's worth faffing with the merge for now if that's the only change.
We'd made some changes to pipelines, but maybe those had already been pulled in. I've just done a git status on our working copy and it says there's nothing else to commit.
If the merge is too much of a pain we'll just delete and re-fork when you get the next release out.
Simon.
Sounds good..
Done as part of #47
See #56 for the code changes.
Most HPC systems have an option to specify expected job execution time, which affects queuing. At the moment, Cluster Flow basically ignores this (SLURM jobs default to two days for example), which hurts our queue performance.
It would be nice if modules are able to specify how long they thing they'll take to run. For SLURM jobs, if this is less than 15 minutes we can get the bonus of specifying
--qos=short
which will make them much faster.