Not being a skilled code reader, may I ask if changeset 24385's "improve the accuracy of FLOPS estimation for GPU apps" applies to anonymous platform?
In the original job runtime estimation whitepaper http://boinc.berkeley.edu/trac/wiki/RuntimeEstimation, it still says, for new hosts and new app_versions without a host_app_version record or insufficient samples, "we use the estimate supplied by the client. ... the current client passes the peak FLOPS." That turned out to be over-optimistic, and in http://boinc.berkeley.edu/trac/changeset/19282 it was changed to "if app_info.xml doesn't specify FLOPS for a GPU app, assume that it runs at CPU peak speed rather than GPU peak speed." Now that GPU computing is more mature, more widespread, and faster, I'd argue that using CPU speed here is over-pessimistic. Could we have some way of deriving a realistic 'first estimate' of effective GPU speed from the client GPU peak speed report and the project-wide experience of GPU application efficiency, please? _______________________________________________ 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.
