[Swift-devel] feature request
Michael Wilde
wilde at mcs.anl.gov
Wed Apr 15 14:36:00 CDT 2009
On 4/15/09 2:22 PM, Mihael Hategan wrote:
> On Wed, 2009-04-15 at 14:05 -0500, Glen Hocky wrote:
>> The problem with the first method was that the number of jobs, i.e.
>> score increased too slowly. In that configuration, i believe that the
>> behavior was 1 or 2 coasters were submitted to a single site and none to
>> the others and then it just stayed that way for a long time.
>
> The behavior you mention is contrary to what should be happening, in
> that all sites should have had 2 swift jobs submitted to.
>
> It is possible that you've made the observation while coasters were not
> working properly on certain sites. The solution to that is not to
> fundamentally re-engineer the way swift submission works, but to make
> coasters run properly on those sites.
I'd like to ask Zhao to do some of this testing, with OOPS on Ranger and
other TG and OSG sites, as prep for getting the tools in broader use in
the OOPS group.
To what extent can we create and maintain a test suite for coasters that
puts it through its paces on a realistic environment. Seems hard, yet
highly desirable, imo.
>
>> Another problem w/ the default configuration was on sites w/ Coasters
>> per node > 1. My experience on ranger with the default parameters is
>> that 2 coasters would start in the queue and only ~6 jobs would run on
>> them, rather than 32. Since our jobs take ~1 hour, this means that for
>> 32 hours of CPU time, I was getting about 6 CPU hours of work. and even
>> after jobs started finishing in that config, the ramp up was too slow
>
> That is indeed a scenario which cannot be addressed by the scheme I
> mentioned. Telling swift that a site has a certain granularity, and 2
> jobs eat the same resources as 16 jobs, is not something we have. Though
> I think the scoring could easily be adapted for that.
>
>
> _______________________________________________
> Swift-devel mailing list
> Swift-devel at ci.uchicago.edu
> http://mail.ci.uchicago.edu/mailman/listinfo/swift-devel
More information about the Swift-devel
mailing list