The hybrid app does about 75% of the work on the CPU, so the run time is
completely dominated by the CPU. The simple solution for tat particular
case would be to schedule based on CPU only, duration correction factor
can easily adjust. I presume the plan class is for 1 CPU and a small
fractional GPU usage, but that doesn't identify how much of the work is
done by each resource. If there were a new value to indicate that balance,
it would be fairly simple to determine the run time contributions.
--
Joe
On 17 Dec 2009 at 15:16, David wrote:
> We'll reduce the scheduler's FLOPS estimate.
> Currently the estimate is (peak GPU FLOPS)/5.
> Does anyone have a suggestion for what it should be?
> Seems like it should reflect both CPU and GPU speed.
>
> -- David
>
> Pappa wrote:
> > The other side effect that has not been fully explored, there were several
> > machines that received over a hundred AP WU's to be errored out. The thought
> > is if it can not determine a proper estimate of run time from the GPU flops
> > and DCF. Only Quota will stop the runaway host.
> >
> > http://setiweb.ssl.berkeley.edu/beta/show_host_detail.php?hostid=42925
> > http://setiweb.ssl.berkeley.edu/beta/results.php?hostid=42925&offset=100&sho
> > w_names=0&state=5
> >
> > http://setiweb.ssl.berkeley.edu/beta/show_host_detail.php?hostid=40712
> > http://setiweb.ssl.berkeley.edu/beta/results.php?hostid=40712&offset=140&sho
> > w_names=0&state=5
> >
> >
> >
> > -----Original Message-----
> > From: boinc_alpha-boun...@ssl.berkeley.edu
> > [mailto:boinc_alpha-boun...@ssl.berkeley.edu] On Behalf Of Stephen Maclagan
> > Sent: Thursday, December 17, 2009 1:23 PM
> > To: boinc_al...@ssl.berkeley.edu
> > Subject: [boinc_alpha] Maximum time Exceeded on Hybrid ATI Astropulse app
> >
> > Raistmer's Hybrid ATI Astropulse app has now made it to Seti Beta as a Stock
> > app, it does only some of it's Calculations on the GPU,
> > with most of it being done on the CPU, some CPU's have been historically
> > poor at doing Astropulse because of their small L2 Caches, ie AMD chips,
> > while the C2D with the Larger caches have been a lot faster, at moment we
> > are starting to seeing some of the AMD's running into maximum time exceeded,
> > because it'll be GPU flops that taken into account when the tasks get
> > aborted,
> > There's also an i7 920 with two HD5800's also running into maximum time
> > exceeded as well, because it has the newest and fastest ATI cards out,
> > while two other i7 920's with lower Spec GPU's can manage to finish the
> > tasks O.K,
> >
> > This was cured in Boinc 6.10.14 with:
> >
> >> - client: if anonymous platform description (app_info.xml) doesn't specify
> > FLOPS for a GPU app, assume that it runs at CPU peak speed rather than GPU
> > peak speed. Better to be conservative, otherwise job might be >aborted due
> > to time limit exceeded.
> >
> > How can it be cured again, now the Hybrid ATI Astropulse app is no longer
> > using an app_info?
> > All the hosts getting aborted tasks are running 6.10.18, and most of the
> > rest are 6.10.18 or newer.
> >
> > See this post for lots of info:
> > http://setiweb.ssl.berkeley.edu/beta/forum_thread.php?id=1601&nowrap=true#38
> > 531
> >
> > Claggy
_______________________________________________
boinc_dev mailing list
boinc_dev@ssl.berkeley.edu
http://lists.ssl.berkeley.edu/mailman/listinfo/boinc_dev
To unsubscribe, visit the above URL and
(near bottom of page) enter your email address.
