On 08-Dec-2010 12:21 PM, Richard Haselgrove wrote: > Now that SETI is back up, details have emerged of a host where the records > appear to indicate that v6.08 (plan_class cuda) is faster than v6.09 > (cuda23). At SETI, this would be unusual - cuda23 is usually significantly > faster than cuda, everything else being equal. > > Two questions. a) How can a user check those figures - I don't know of any > logging options which could help.
It's available on the web; see e.g. http://setiathome.berkeley.edu/host_app_versions.php?hostid=5116976 which indeed says that cuda is faster than cuda23 for this host b) If the server has latched onto a > sub-optimal application version - perhaps one application's records were > skewed by a bad batch of tasks, or a driver problem - is there any way the > user can trigger a re-evaluation, perhaps by resetting the historical > averages? Possible approaches: - clear the (host, app version) DB records when the GPU HW or SW change, as you suggest. This wouldn't address the problem of an estimate that's off because of an unusual batch of tasks. - multiply by a random factor when comparing the performance of different app versions. So, in this case, every so often the scheduler would use cuda23 even though it thinks cuda is faster, and eventually cuda23 would (possibly) exceed cuda. The random factor could be normal with mean 1 and stddev .1 or so -- David _______________________________________________ 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.
