Lynn W. Taylor wrote: [...] > Paul, I know you're mad because you're sure that your benchmark idea is > so obviously correct that you can't imagine why it would not be > instantly adopted. > > In reality, for many projects the "benchmark" would change based on what > standard work unit was chosen. Just look at all of the discussion about > VLAR and VHAR work. It's especially bad for some angle ranges and CUDA > cards. [...]
From some of Paul's voluminous detailed ramblings, I must agree! In brief, there usually is a good story in there ;-) The benchmark vs NIST-style calibration has long been thrashed out. The real problem at the moment is that the credits system is crediting for aspects that are not even proportional to some of the work that is supposedly being credited for. All from a near meaningless benchmark that is only crudely non-linearly proportional to what s...@h does. So... Avoiding repeating a long discussion... In brief... You don't "benchmark" an (un)representative WU or benchmark routine on a participants host, all at vast expense of wasted cycles to still then get the wrong benchmark... Instead, as in all good Science, you *calibrate* against a "Golden Standard". A simple example for Boinc would be an actual physical computer at the project lab or wherever, provided that it can run through all types of WUs, reasonably quickly, and survive the lifespan of the project. That is then used to calibrate *live* WUs across all participants, in a hierarchical way similar to what is done with NIST standards. The only requirement is that there is enough WU redundancy to allow enough of an amplification factor for the hierarchical calibration to work. The WU redundancy could be made adaptive so that new hosts are paired up against another calibrated host to gain a level of calibration on their very first WU. (I'm sure Paul can elaborate on suitable algorithms if asked.) In that scheme, *there are no "test" WUs* ! ALL the WUs are live work. Regards, Martin -- -------------------- Martin Lomas m_boincdev ml1 co uk.ddSPAM.dd -------------------- _______________________________________________ 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.
