> With the real-world measurements and calibration, we can then also > measurably say how well particular hardware performs for the WUs as > compared to the Etalon system. That could well run riot with the Boinc > stats sites if they can honestly post up the "top 10" types of systems > for performance!
Good point even be considered alone! Having such knowledge further optimization of BOINC supercomputer possible (or, at least some subsets of it - participant's farms). If, by simple direct comparison of credit per hour for 2 different apps belonging to the same or to different projects one participates will be possible to get relative performance measure what app better fits to what hardware one could change project shares or app sets running on different PCs in farm and recive additional increase in overall farm performance. Same approach being applied on project level (for projects with few apps send particular type of work preferably to hardware that does this particular type of work better among other hardware still capable to do work) will increase BOINC's hyper-computer performance as whole. Currently, w/o real inter-project credit parity, such measurements inherently contain systematic errors. > > For maintaining the calibration with newer "Etalon" systems, all that > should change would be the effect of the 'imponderables' which the > Cobblestones at present don't measure in the first place. It will mean > that the participants hosts will be realigned depending on how well > their 'imponderables' compare to the new reference machine architecture. And how well particular hardware/software combo performs realtively to good known standart. Valuable info for wise chosing software/hardware combos. > Hopefully, the 'credits' get updated to credit for more than just > Cobblestones so that they directly include the various 'imponderables' > that are not measured nor directly credited for at present. That is, we keeping "currency" aspect, keeping competition for credits (as attractive factor for some participants) + turning credits in good performance measure tool that can be used not only for bragging but for performance optimization as well. > Even if that isn't done, keeping to a middle-of-the-road Etalon just > means that the 'imponderables' part of the calibration evolves with the > replacement hardware and so should never be more than a few CPU > generations out of step with the rest of the world. > > If a significant change due to 'imponderables' were to be noticed, then > we should be directly crediting for that in the first place in any case! Yes. Currently very paradoxal situations occur. Participant can compare benchmarks for his (or some other) Core2 and AMD-based PCs, it can give pretty close values then it will try to run AstroPulse for example on both and recive pretty poor crediting on AMD host. Natural question arise - why poor credit/hour if benchmarks were pretty well ? Some hardware better fits to one work and some to another. Doing benchmarks (for sake of crediting) on participant's hardware obscure this fact and creates such strange situations. >> 1) The conversion factor "c" is going to be different for every project, >> because the instruction mix is different for every project, and the >> importance of memory performance, cache size, and factors that don't >> affect whetstone or dhrystone numbers but do affect crunching time. A >> single in-lab reference machine has to crunch all projects. Some algorithms just doesn't fit well in GPU architecture. But doing conversion factor different can obscure this fact. It could be step back IMO. If task can't be parallelized enough to make good use of GPU than GPU not very needed for such project and not need to keep pay rate for GPU work from such project the same as for project fully utilizing GPU shaders. >> ... and it only really matters to the volunteers. It matters to the >> BOINC developers because they have to encourage participation, and >> credit does that. It doesn't matter much to the projects themselves, >> probably not enough to set up a reference-fleet of their own to grant >> accurate credit. It could matter for projects that care about their hyper-computer performance. _______________________________________________ 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.
