The BOINC message board thread needs to be read in parallel with the matching NumberFields thread http://numberfields.asu.edu/NumberFields/forum_thread.php?id=209
In the project thread, the project administrator confirms that the project is using 'credit from runtime' (the task lengths are indeterminate). Given that, the credit formula is (from l461 of validator.cpp) credit = result.flops_estimate * runtime * COBBLESTONE_SCALE; >From what people have told me, the variable result.flops_estimate in that >context contains a value closely tied to the Whetstone benchmark for the host. >Numberfields is, I believe, a largely integer math project. In that case, >could the disparity in credit boil down to the relative integer and floating >point strengths of AMD and Intel CPUs? In particular, if AMD has a low runtime (because its integer math unit is strong), and has a low result.flops_estimate (because its FP unit is weak), credit from runtime suffers on both parameters. >________________________________ > From: Jord van der Elst <[email protected]> >To: David Anderson <[email protected]> >Cc: BOINC Dev Mailing List <[email protected]> >Sent: Friday, August 29, 2014 4:45 PM >Subject: [boinc_dev] Credit AMD CPU vs Intel CPU. > > >Hi David, > >In https://boinc.berkeley.edu/dev/forum_thread.php?id=9569 user >"Grandpa" reports that he has found a big discrepancy between credit >given out to AMD CPUs versus credit given out to Intel CPUs, at the >Numberfields project. > >I'll quote his last post on this: > >-------- >AMD Opteron(TM) Processor 6276 [Family 21 Model 1 Stepping 2] >(64 processors) >Total runtime 100 WU's 1810010.65 Total points 100 WU's 26128.8 Grandma 6276 >AMD >Avg runtime per WU sec. 18100.1065 avg points per WU 261.288 3042Mhz >Avg runtime per WU min. 301.6684416667 >Avg runtime per WU hr. 5.0278073611 avg points per hr runtime 51.9685781959 > >Genuine Intel(R) CPU @ 2.70GHz [Family 6 Model 45 Stepping 5] >(64 processors) >Total runtime 100 WU's 1839477.54 Total points 100 WU's 38003.75 Musky >4650 Intel >Avg runtime per WU sec. 18394.7754 avg points per WU 380.0375 3134Mhz >Avg runtime per WU min. 306.57959 >Avg runtime per WU hr. 5.1096598333 avg points per hr runtime 74.3762818653 > >OK but it still does not change the fact that this type of credit >system has a pretty big flaw in it. According to the developer at >Numberfields each of the machines above are doing the same amount of >work and lets say that each WU is worth 1 FLOPS since the numbers were >averaged over 100 WU's each and all things being equal those 2 >machines should have received roughly the same amount of credit with a >slight advantage going to the AMD, but in actuality AMD is receiving >30% less. > >I do not really see this as being a fair credit system to the AMD >users when it come to certain types of work. > >AMD 100 WU's = 100 FLOPS = 502.78 hrs of runtime = 26706.79 points of >credit for 100 theoretical FLOPS > >Intel 100 WU's = 100 FLOPS = 510.96 hrs of runtime = 38003.74 points >of credit for 100 theoretical FLOPS > >So in plain simple terms AMD gets less credit for doing the same >amount of work as Intel does. I am just pointing out a pretty big flaw >in the current credit system, from everything I have read and been >told this system was set up to promote equality between all work done >and to try and discourage cheating, It appears to me that it may have >missed it's mark when in the equality field and they may need to go >back to the drawing board and try and fix the problem or at least let >people know that the credit system has a problem with some projects. >-------- > >I seem to remember that Seti has a similar problem, but the finer >details escape me. > >Perhaps that (more) knowledgeable people can explain the >discrepancy/difference? > >Thanks, > >-- Jord van der Elst. >_______________________________________________ >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.
