[Swift-devel] fast-failing jobs
Mihael Hategan
hategan at mcs.anl.gov
Sun Apr 13 11:22:40 CDT 2008
On Sun, 2008-04-13 at 16:19 +0000, Ben Clifford wrote:
> On Sat, 12 Apr 2008, Mihael Hategan wrote:
>
> > That's pretty much what a low score does if there's throttling based on
> > score. Perhaps our solution is to have a low job throttle and a higher
> > score range (i.e. T=1000 instead of 100).
>
> The present scoring system won't ever go below 2 jobs per site, so pretty
> much whatever the parameters are tweaked to, a fast-fail site will eat 2
> jobs per fast-fail cycle.
That being one thing that probably should be changed.
>
> > That or we could enforce a submission rate (j/s) based on score.
>
> That is perhaps better.
>
> It would make lower scores more punitive than at the moment, which may be
> a problem given the way that in certain other failure modes the score gets
> reduced catastrophically. (eg a transient problem where all jobs fail that
> are in progress, with a large number of jobs in progress - this was why I
> put that lower bound in on the scores - to prevent the score actually
> getting really low)
>
More information about the Swift-devel
mailing list