[Beowulf] scheduler policy design
hahn at mcmaster.ca
Thu Apr 26 09:34:01 PDT 2007
> We do advance reservation with backfill, so short jobs will on average
> start sooner than long jobs. We find that this is sufficient
> motivation for the users to at least try to set reasonable walltime
have you examined the accuracy of your user's estimates,
both underruns and overruns? the latter will presumably create
some idle cycles, but the latter (if you kill past reservation)
will potentially waste quite large amounts of work.
we've looked at the literature on reservations, and the few studies
that mention user accuracy seem to agree that the problem is dire.
More information about the Beowulf