> ...argue for the event based, then time based then both ...
I've followed this thread with interest, so here's my small contribution.
Obviously, if event based could be guaranteed to occur whenever needed,
the safety net of time based would not be needed. If someone had a few
months to do extensive testing and produce a super stable version...
So the basic issue is just how paranoid the time based safety net should
be. I tend to agree that the one minute setting is excessive nervousness.
Changing it to ten minutes involves adding a single '0' character.
The interaction with rr_simulation could be improved if further refinement
seems justified. That is, if the simulation indicates that all work will
complete within a day and the nearest computation deadline is ten days
away, it might make sense to put the safety net away for at least a day
or so. There's no output from an rr_sim which could be used for that
purpose, but adding one would be fairly simple.
My feeling is if a host has a very large number of tasks cached, the
time spent running rr_sims might actually be measurably reducing the time
given to applications. I doubt it's among the worst offenders, though.
Overall, I agree with those who want to focus on why some scheduling actions
seem not to make sense. That's the part which is critical.
--
Joe
_______________________________________________
boinc_dev mailing list
[email protected]
http://lists.ssl.berkeley.edu/mailman/listinfo/boinc_dev
To unsubscribe, visit the above URL and
(near bottom of page) enter your email address.
