This is a known problem with the design of the scheduler. The "conservative" estimates used for versions with fewer than 10 normal results are usually worse than the estimates for any version with more than 10 results even when random factors are added. Eventually it fixes itself. I've been considering changes that could fix it, but that hasn't been implemented because they might interfere with other aspects of the scheduler.
On Thu, Jul 10, 2014 at 12:09 PM, Raistmer the Sorcerer <[email protected]> wrote: > The same situation I saw when SETI released few CUDA app versions. Host > was "locked" on suboptimal one just because it collects 10 eligible valid > results before others. > > > > Thu, 10 Jul 2014 19:24:53 +0100 от Stephen Maclagan < > [email protected]>: > >I reported this on Seti Beta a few days ago, no response: > > > > > http://setiweb.ssl.berkeley.edu/beta/forum_thread.php?id=2182&postid=51497 > > > >Basically Seti Beta has new Astropulse v7 apps, both Stock based and > Optimised based, > >for Linux the Stock and Optimised apps were released together, my C2D > T8100 has managed to get a mix of work from the four app_versions, > >none of which have got to the their 11 validations yet: > > > >http://setiweb.ssl.berkeley.edu/beta/host_app_versions.php?hostid=68093 > > > >For the windows Astropulse v7 apps, there was a four day difference > between the slow Stock app being released and the fast optimised app being > released, > >meaning my i7-2600K was already a good way through crunching the slow > Stock apps before the Optimised apps were released, > >it did manage to receive a couple of token SSE and SSE2 tasks, they are > up six time faster, > > > >http://setiweb.ssl.berkeley.edu/beta/host_app_versions.php?hostid=45274 > > > >http://setiweb.ssl.berkeley.edu/beta/results.php?hostid=45274 > > > >The i7-2600K now has 19 validations for windows_intelx86 app, one for the > windows_x86_64 (sse2) app, and three for the windows_intelx86 (sse) app, > >Inspite of these apps being a huge amount faster, the scheduler would > rather send work for the slow app with greater than 11 validations, > >than a faster app with less validations, > > > >Shouldn't the scheduler try and get each app version it's 11 validations > first before deciding which app is fastest? > > > >Claggy > > > >_______________________________________________ > >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. > > _______________________________________________ > 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. > > _______________________________________________ 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.
