I checked in a change that does this (it involved commenting out the code that tries to maintain some GPU work for all projects) -- David
[email protected] wrote: > There is a mismatch between the CUDA resource scheduler and the CPU work > fetch algorithm. The CUDA resource scheduler is FIFO with occasional > breaks for EDF. The CPU resource scheduler is round robin. I believe that > both use a work fetch that tries to fetch from as many different projects > as possible - even if the cache is over filled in some cases. While this > is very good for a round robin resource scheduler, it is a bit of a > mis-match for a FIFO resource scheduler. The FIFO resource scheduler will > not get around to some of the work for at least the size of the cache time. > > I would propose that the FIFO resource scheduler be matched with a work > fetch that stops trying to fetch work from anywhere as soon as the queue is > filled for all resources. > > jm7 > > _______________________________________________ > boinc_sched mailing list > [email protected] > http://lists.ssl.berkeley.edu/mailman/listinfo/boinc_sched _______________________________________________ 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.
