I'm trying to figure out how get a plan class chosen reliably in a plan_class_spec.xml plan class. It seems easier for me to add a <without_opencl> boolean to the plan class without having to mess around with FLOPS estimates.
There are several issue that I haven't quite figured out in version selection. One is that versions seem to be selected based on shortest run time rather highest average processing rate. In SETI@home, where run times are quite variable, this seems to often result in the wrong version being selected, even with outlier exclusion. On Tue, Jan 29, 2013 at 4:27 PM, David Anderson <[email protected]> wrote: > The project should adjust its app_plan() function so that > the estimated FLOPS for the OpenCL version > is higher than that for the CAL version. > -- David > > > On 29-Jan-2013 4:01 PM, Raistmer the Sorcerer wrote: >> >> SETI@home has both OpenCL ATi and Brook+ (CAL) apps for one of its >> subprojects. Currently CAL app distributed to all CAL-capable ATi >> devices, >> including OpenCL capable ones. As was shown with CUDA apps on SETI beta >> BOINC >> can't select fastest app if few available. CAL app much slower than OpenCL >> app. So, feature needed to allow server to send CAL app only to CAL-only >> capable hosts (not send if host supports OpenCL too) and OpenCL app for >> OpenCL+CAL/OpenCL hosts. AFAIK currently it''s not possible. >> >> _______________________________________________ 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.
