Then it looks like both of the GPU projects I have my computers participate in have done a rather poor job of that, and need some extra help doing it properly.
One consistently sets the deadlines a few days AFTER the time they start issuing additional copies of the workunit to other users. For that one, the combination of a short queue time (not a problem on that computer) and a fairly fast GPU allows me to finish all their workunits before they start sending out additional copies. For this one, it would help to allow a setting specific to that project to tell BOINC to make the first effort to report completed workunits perhaps 5 minutes after uploading the last file from a completed workunit. The other consistently sends me workunits with the initial estimated time to completion around 20 times what it actually takes, and around twice as many workunits as that computer (with a rather slow GPU, and also with the Tthrottle add-on to slow it down when it would otherwise overheat) can complete by the deadline. This computer is a laptop, and is unlikely to allow upgrading to a faster GPU, unless a new generation of GPUs allows running perhaps twice as fast without using any more power. Another idea to allow that project to work better: Allow the USER, not just the project, to set a limit on the number of GPU workunits in the queue for that project, possibly with one more allowed if the total estimated remaining time to completion for all GPU workunits for all projects drops below a certain value (such as 1 hour). I'd expect that exception to also make it friendlier to share the same GPU among multiple BOINC projects with GPU workunits. On 20-Apr-2011,David Anderson wrote: > Projects that want fast turnaround time for GPU jobs > either assign short deadlines for those jobs, > or limit the # of GPU jobs in progress on a host. > Queue length doesn't impact either of these. > -- David > > On 19-Apr-2011 9:49 PM, robert miles wrote: >> Could you make the queue length for GPU workunits independent from the queue >> length for CPU workunits? >> >> On my laptop, I currently need to make the queue length at least 3 days in >>order >> to make it download CPU workunits at all. However, most projects that have >>gone >> to the effort of creating GPU workunits at all want their workunits not to >> remain in queues that long, and therefore a shorter queue length for GPU >> workunits would be appropriate ? perhaps an hour or less. _______________________________________________ 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.
