[Clusterusers] Job Sorting Schemes

[Jaime Davila]

If I'm reading this text right, I don't think those different schemes 
would make much difference, given all jobs with the same priority, since 
what we have right now is a mix of tasks that take very long and tasks 
that take very little running together in the same system, and their 
parent jobs all want to have the same number of children spawn (100).

FWIW, by generation 6 I can tell that what I started doing works. How 
well, that will take time to figure out, but speed is not terribly 
critical for me right now. I'll always be happy with more CPU cycles, 
but if other stuff is running on a tight deadline, I'll live with what I 
have.

Thanks for all your work,

Jaime


On 06/28/2013 01:55 PM, Wm. Josiah Erikson wrote:
> Just in case people are interested, we could also try playing with 
> this setting. We might want to try P+ATCL. We are currently using P+FIFO:
>
> #
>   # Job Sorting Schemes for Blade Assignment
>   #
>   # "P+FIFO" -- this is the default scheme for assigning available
>   # blades to jobs, it always picks the job with the highest priority
>   # value, and if there are several jobs with the same priority then
>   # it selects the one that was spooled first.
>   #
>   # "P+RR" -- this mode schedules jobs of the same priority using a
>   # simple "Round-Robin" approach:  jobs in the same priority block
>   # are visited in a circular fashion as blades become available.
>   # This mode can allow several jobs with many ready tasks to
>   # progress simultaneously, favoring jobs with long-running tasks,
>   # although it is not as predictable as the strict "P+FIFO" mode.
>   #
>   # "P+ATCL" -- Active Task Count Leveling, this mode also first
>   # sorts jobs by strict priority, then within a group of jobs with
>   # the same priority it prefers to assign available blades to jobs
>   # with the fewest active tasks.  This mode tends to allocate roughly
>   # the same number of blades to each job, while favoring older jobs
>   # over newly spooled ones.  Given roughly equal numbers of blades,
>   # jobs with short-running tasks will finish sooner than jobs with
>   # long-running tasks, under this scheme.
>   #
>   # NOTE: Any scheduling mode other than P+FIFO will incur
>   # extra computational expense during blade assignment, and
>   # will therefore decrease system throughput by some amount.
>   # The magnitude of the observed effect will vary depending
>   # on overall job load, farm size, task mix, and engine host.
>   #
>   "JobSchedulingMode": "P+FIFO",
>
> -- 
> Wm. Josiah Erikson
> Assistant Director of IT, Infrastructure Group
> System Administrator, School of CS
> Hampshire College
> Amherst, MA 01002
> (413) 559-6091
>
>
> _______________________________________________
> Clusterusers mailing list
> Clusterusers at lists.hampshire.edu
> https://lists.hampshire.edu/mailman/listinfo/clusterusers

-- 
******************************************************
Jaime J. Dávila
Associate Professor of Computer Science
Hampshire College
jdavila at hampshire dot edu
http://helios.hampshire.edu/jdavila
*******************************************************

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.hampshire.edu/pipermail/clusterusers/attachments/20130628/c9394da2/attachment.html>