On May 30, 2008, at 7:56 AM, Sean McGrath - Sun Microsystems Ireland wrote:
> eric kustarz stated: > < > > < > Currently both tools gives different view of io. While filebench > < > simulates real workload, it cannot show what vdbench shows. > < > E.g. we were doing some hardware array tests and it turned out > (using > < > vdbench) that the array works in a really strange way: sometimes > its > < > IOPS jumps very high, sometimes it lows very much. There were > < > a few other tests where vdbench shows how IOPS behaves in each > < > second of workload. I am affraid filebench _currently_ cannot > give us > < > the same data. > < > < Good point. > < > < We actually have this via Xanadu, though i just tried it and it > looks > < like Xanadu is not working. > < > < Also, at the UCSC benchmarking conference last monday we kicked > around > < the idea of showing a distribution of results instead of just > averages. > > Yes, averages tell only part of the story. Agreed. > > > as an example, libmicro has some nice statistics that it gives with > its metric > results. Could the methods be used within filebench ? Possibly, its something we're discussing so we'll take input (like yours)... I like the distribution output (perhaps similar to Dtrace's quantize) better though. eric > > > eg: > > # STATISTICS usecs/call (raw) usecs/call (outliers > removed) > # min 0.00448 0.00448 > # max 0.00623 0.00459 > # mean 0.00453 0.00451 > # median 0.00453 0.00453 > # stddev 0.00013 0.00003 > # standard error 0.00001 0.00000 > # 99% confidence level 0.00002 0.00000 > # skew 11.81224 0.45940 > # kurtosis 153.61917 -0.55683 > # time correlation -0.00000 -0.00000 > # > # elasped time 0.01943 > # number of samples 196 > # number of outliers 6 > # getnsecs overhead 288 > > < > < eric > < _______________________________________________ > < perf-discuss mailing list > < perf-discuss@opensolaris.org > > -- > Sean. > . _______________________________________________ perf-discuss mailing list perf-discuss@opensolaris.org
