From: Andrew Morton <akpm@os...>  20030822 21:02:28

Steve, thanks for summarizing the results in this manner. Frankly I wouldn't have noticed *any* of these regressions had you not performed this analysis  there's a lot of valueadd on top of the raw numbers here. I'd be suspecting the CPU scheduler changes for most things, but perhaps not the dbenchonext3 problem. RAMonly dbench shouldn't regress like that; I can take a look at it. volanomark and specjbb are harder for me to analyse. Could someone at your end please do a quick A/B test? Grab a current Linus tree (2.6.0test3/bk8 will do) and test volanomark or specjbb with and without http://www.zip.com.au/~akpm/linux/patches/schedrollup.patch I wouldn't do an exhaustive test: just enough to be able to say "yup, that's the culprit" or "no, it's something else". Thanks. Steven Pratt <slpratt@...> wrote: > > Mark Peloquin wrote: > > > Been awhile since results where posted, therefore this is a little long. > > > > > > Nightly Regression Summary for 2.6.0test3 vs 2.6.0test3mm3 > > > > Benchmark Pass/Fail Improvements Regressions > > Results Results Summary > >     > >    > > dbench.ext2 P N N 2.6.0test3 > > 2.6.0test3mm3 report > > dbench.ext3 P N Y 2.6.0test3 > > 2.6.0test3mm3 report > > The ext3 dbench regression is very significant for multi threaded 193 > > 118. Looks like this regression first showed up in mm1 and does not > exist in any of the bk trees. > > http://ltcperf.ncsa.uiuc.edu/data/historygraphs/dbench.ext3.throughput.plot.16.png > > > > > volanomark P N Y 2.6.0test3 > > 2.6.0test3mm3 report > > Volanomark is significant as well. 10% drop in mm tree. This one also > appeared to show up in mm1 although it was a 14% drop then so mm3 > actually looks a little better. There were build errors on mm2 run so I > don't have that data at this time. > Following link illustrates the drop in mm tree for volanomark. > > http://ltcperf.ncsa.uiuc.edu/data/historygraphs/volanomark.throughput.plot.1.png > > > SpecJBB2000 for high warehouses also took a bit hit. Probably the same > root cause as volanomark. > Here is the history plot for the 19 warehouse run. > > http://ltcperf.ncsa.uiuc.edu/data/historygraphs/specjbb.results.avg.plot.19.png > > Huge spike in idle time. > http://ltcperf.ncsa.uiuc.edu/data/historygraphs/specjbb.utilization.idle.avg.plot.19.png > > > > > http://ltcperf.ncsa.uiuc.edu/data/2.6.0test3mm3/2.6.0test3vs2.6.0test3mm3/ > > > > > Steve > 
From: Mark Peloquin <peloquin@au...>  20030821 22:41:51

Been awhile since results where posted, therefore this is a little long. Nightly Regression Summary for 2.6.0test3 vs 2.6.0test3mm3 Benchmark Pass/Fail Improvements Regressions Results Results Summary        dbench.ext2 P N N 2.6.0test3 2.6.0test3mm3 report dbench.ext3 P N Y 2.6.0test3 2.6.0test3mm3 report dbench.jfs P N N 2.6.0test3 2.6.0test3mm3 report dbench.reiser P N N 2.6.0test3 2.6.0test3mm3 report dbench.xfs P N N 2.6.0test3 2.6.0test3mm3 report kernbench P N N 2.6.0test3 2.6.0test3mm3 report lmbench P Y Y 2.6.0test3 2.6.0test3mm3 report rawiobench P Y N 2.6.0test3 2.6.0test3mm3 report specjbb P Y Y 2.6.0test3 2.6.0test3mm3 report specsdet P N Y 2.6.0test3 2.6.0test3mm3 report tbench P Y Y 2.6.0test3 2.6.0test3mm3 report tiobench.ext2 P Y Y 2.6.0test3 2.6.0test3mm3 report tiobench.ext3 P Y Y 2.6.0test3 2.6.0test3mm3 report tiobench.jfs P N Y 2.6.0test3 2.6.0test3mm3 report tiobench.reiser P Y Y 2.6.0test3 2.6.0test3mm3 report tiobench.xfs P Y Y 2.6.0test3 2.6.0test3mm3 report volanomark P N Y 2.6.0test3 2.6.0test3mm3 report http://ltcperf.ncsa.uiuc.edu/data/2.6.0test3mm3/2.6.0test3vs2.6.0test3mm3/ Nightly Regression Summary for 2.6.0test3 vs 2.6.0test3mm1 Benchmark Pass/Fail Improvements Regressions Results Results Summary        dbench.ext2 P N N 2.6.0test3 2.6.0test3mm1 report dbench.ext3 P N Y 2.6.0test3 2.6.0test3mm1 report dbench.jfs P N N 2.6.0test3 2.6.0test3mm1 report dbench.reiser P N N 2.6.0test3 2.6.0test3mm1 report dbench.xfs P N N 2.6.0test3 2.6.0test3mm1 report kernbench P N N 2.6.0test3 2.6.0test3mm1 report lmbench P Y Y 2.6.0test3 2.6.0test3mm1 report rawiobench P Y Y 2.6.0test3 2.6.0test3mm1 report specjbb P Y Y 2.6.0test3 2.6.0test3mm1 report specsdet P N Y 2.6.0test3 2.6.0test3mm1 report tbench P Y Y 2.6.0test3 2.6.0test3mm1 report tiobench.ext2 P N N 2.6.0test3 2.6.0test3mm1 report tiobench.ext3 P Y Y 2.6.0test3 2.6.0test3mm1 report tiobench.jfs P Y Y 2.6.0test3 2.6.0test3mm1 report tiobench.reiser P Y Y 2.6.0test3 2.6.0test3mm1 report tiobench.xfs P Y N 2.6.0test3 2.6.0test3mm1 report volanomark P N Y 2.6.0test3 2.6.0test3mm1 report http://ltcperf.ncsa.uiuc.edu/data/2.6.0test3mm1/2.6.0test3vs2.6.0test3mm1/ Nightly Regression Summary for 2.6.0test3 vs 2.6.0test3bk8 Benchmark Pass/Fail Improvements Regressions Results Results Summary        dbench.ext2 P N N 2.6.0test3 2.6.0test3bk8 report dbench.ext3 P N N 2.6.0test3 2.6.0test3bk8 report dbench.jfs P N N 2.6.0test3 2.6.0test3bk8 report dbench.reiser P N N 2.6.0test3 2.6.0test3bk8 report dbench.xfs P N N 2.6.0test3 2.6.0test3bk8 report kernbench P N N 2.6.0test3 2.6.0test3bk8 report lmbench P Y Y 2.6.0test3 2.6.0test3bk8 report rawiobench P Y Y 2.6.0test3 2.6.0test3bk8 report specjbb P Y Y 2.6.0test3 2.6.0test3bk8 report specsdet P N N 2.6.0test3 2.6.0test3bk8 report tbench P N Y 2.6.0test3 2.6.0test3bk8 report tiobench.ext2 P N Y 2.6.0test3 2.6.0test3bk8 report tiobench.ext3 P Y N 2.6.0test3 2.6.0test3bk8 report tiobench.jfs P N Y 2.6.0test3 2.6.0test3bk8 report tiobench.reiser P N Y 2.6.0test3 2.6.0test3bk8 report tiobench.xfs P N Y 2.6.0test3 2.6.0test3bk8 report volanomark P N N 2.6.0test3 2.6.0test3bk8 report http://ltcperf.ncsa.uiuc.edu/data/2.6.0test3bk8/2.6.0test3vs2.6.0test3bk8/ Nightly Regression Summary for 2.6.0test3bk5 vs 2.6.0test3bk6 Benchmark Pass/Fail Improvements Regressions Results Results Summary        dbench.ext2 P N N 2.6.0test3bk5 2.6.0test3bk6 report dbench.ext3 P N N 2.6.0test3bk5 2.6.0test3bk6 report dbench.jfs P N N 2.6.0test3bk5 2.6.0test3bk6 report dbench.reiser P N N 2.6.0test3bk5 2.6.0test3bk6 report dbench.xfs P N Y 2.6.0test3bk5 2.6.0test3bk6 report kernbench P N Y 2.6.0test3bk5 2.6.0test3bk6 report lmbench P Y Y 2.6.0test3bk5 2.6.0test3bk6 report rawiobench P Y N 2.6.0test3bk5 2.6.0test3bk6 report specjbb P Y N 2.6.0test3bk5 2.6.0test3bk6 report specsdet P N N 2.6.0test3bk5 2.6.0test3bk6 report tbench P N Y 2.6.0test3bk5 2.6.0test3bk6 report tiobench.ext2 P N N 2.6.0test3bk5 2.6.0test3bk6 report tiobench.ext3 P N N 2.6.0test3bk5 2.6.0test3bk6 report tiobench.jfs P N N 2.6.0test3bk5 2.6.0test3bk6 report tiobench.reiser P Y Y 2.6.0test3bk5 2.6.0test3bk6 report tiobench.xfs P N N 2.6.0test3bk5 2.6.0test3bk6 report volanomark P N N 2.6.0test3bk5 2.6.0test3bk6 report http://ltcperf.ncsa.uiuc.edu/data/2.6.0test3bk6/2.6.0test3bk5vs2.6.0test3bk6/ Nightly Regression Summary for 2.6.0test3 vs 2.6.0test3bk6 Benchmark Pass/Fail Improvements Regressions Results Results Summary        dbench.ext2 P N N 2.6.0test3 2.6.0test3bk6 report dbench.ext3 P N N 2.6.0test3 2.6.0test3bk6 report dbench.jfs P N N 2.6.0test3 2.6.0test3bk6 report dbench.reiser P N N 2.6.0test3 2.6.0test3bk6 report dbench.xfs P N N 2.6.0test3 2.6.0test3bk6 report kernbench P N N 2.6.0test3 2.6.0test3bk6 report lmbench P Y Y 2.6.0test3 2.6.0test3bk6 report rawiobench P Y N 2.6.0test3 2.6.0test3bk6 report specjbb P Y Y 2.6.0test3 2.6.0test3bk6 report specsdet P Y N 2.6.0test3 2.6.0test3bk6 report tbench P N N 2.6.0test3 2.6.0test3bk6 report tiobench.ext2 P Y N 2.6.0test3 2.6.0test3bk6 report tiobench.ext3 P Y N 2.6.0test3 2.6.0test3bk6 report tiobench.jfs P N Y 2.6.0test3 2.6.0test3bk6 report tiobench.reiser P N N 2.6.0test3 2.6.0test3bk6 report tiobench.xfs P N Y 2.6.0test3 2.6.0test3bk6 report volanomark P N N 2.6.0test3 2.6.0test3bk6 report http://ltcperf.ncsa.uiuc.edu/data/2.6.0test3bk6/2.6.0test3vs2.6.0test3bk6/ Nightly Regression Summary for 2.6.0test3 vs 2.6.0test3bk5 Benchmark Pass/Fail Improvements Regressions Results Results Summary        dbench.ext2 P N N 2.6.0test3 2.6.0test3bk5 report dbench.ext3 P N N 2.6.0test3 2.6.0test3bk5 report dbench.jfs P N N 2.6.0test3 2.6.0test3bk5 report dbench.reiser P N N 2.6.0test3 2.6.0test3bk5 report dbench.xfs P N N 2.6.0test3 2.6.0test3bk5 report kernbench P Y N 2.6.0test3 2.6.0test3bk5 report lmbench P Y Y 2.6.0test3 2.6.0test3bk5 report rawiobench P Y Y 2.6.0test3 2.6.0test3bk5 report specjbb P Y Y 2.6.0test3 2.6.0test3bk5 report specsdet P N N 2.6.0test3 2.6.0test3bk5 report tbench P Y N 2.6.0test3 2.6.0test3bk5 report tiobench.ext2 P Y N 2.6.0test3 2.6.0test3bk5 report tiobench.ext3 P Y N 2.6.0test3 2.6.0test3bk5 report tiobench.jfs P N Y 2.6.0test3 2.6.0test3bk5 report tiobench.reiser P N N 2.6.0test3 2.6.0test3bk5 report tiobench.xfs P N Y 2.6.0test3 2.6.0test3bk5 report volanomark P N N 2.6.0test3 2.6.0test3bk5 report http://ltcperf.ncsa.uiuc.edu/data/2.6.0test3bk5/2.6.0test3vs2.6.0test3bk5/ Nightly Regression Summary for 2.6.0test3bk2 vs 2.6.0test3bk3 Benchmark Pass/Fail Improvements Regressions Results Results Summary        dbench.ext2 P N N 2.6.0test3bk2 2.6.0test3bk3 report dbench.ext3 P N N 2.6.0test3bk2 2.6.0test3bk3 report dbench.jfs P N N 2.6.0test3bk2 2.6.0test3bk3 report dbench.reiser P N N 2.6.0test3bk2 2.6.0test3bk3 report dbench.xfs P N N 2.6.0test3bk2 2.6.0test3bk3 report kernbench P Y N 2.6.0test3bk2 2.6.0test3bk3 report lmbench P Y Y 2.6.0test3bk2 2.6.0test3bk3 report rawiobench P Y Y 2.6.0test3bk2 2.6.0test3bk3 report specjbb P Y N 2.6.0test3bk2 2.6.0test3bk3 report specsdet P N N 2.6.0test3bk2 2.6.0test3bk3 report tbench P Y Y 2.6.0test3bk2 2.6.0test3bk3 report tiobench.ext2 P Y N 2.6.0test3bk2 2.6.0test3bk3 report tiobench.jfs P N N 2.6.0test3bk2 2.6.0test3bk3 report tiobench.reiser P Y N 2.6.0test3bk2 2.6.0test3bk3 report tiobench.xfs P Y Y 2.6.0test3bk2 2.6.0test3bk3 report volanomark P N N 2.6.0test3bk2 2.6.0test3bk3 report http://ltcperf.ncsa.uiuc.edu/data/2.6.0test3bk3/2.6.0test3bk2vs2.6.0test3bk3/ Nightly Regression Summary for 2.6.0test3 vs 2.6.0test3bk3 Benchmark Pass/Fail Improvements Regressions Results Results Summary        dbench.ext2 P N N 2.6.0test3 2.6.0test3bk3 report dbench.ext3 P N N 2.6.0test3 2.6.0test3bk3 report dbench.jfs P N N 2.6.0test3 2.6.0test3bk3 report dbench.reiser P N N 2.6.0test3 2.6.0test3bk3 report dbench.xfs P N N 2.6.0test3 2.6.0test3bk3 report kernbench P Y N 2.6.0test3 2.6.0test3bk3 report lmbench P Y Y 2.6.0test3 2.6.0test3bk3 report rawiobench P Y Y 2.6.0test3 2.6.0test3bk3 report specjbb P Y Y 2.6.0test3 2.6.0test3bk3 report specsdet P N N 2.6.0test3 2.6.0test3bk3 report tbench P Y Y 2.6.0test3 2.6.0test3bk3 report tiobench.ext2 P Y N 2.6.0test3 2.6.0test3bk3 report tiobench.jfs P N N 2.6.0test3 2.6.0test3bk3 report tiobench.reiser P N Y 2.6.0test3 2.6.0test3bk3 report tiobench.xfs P Y Y 2.6.0test3 2.6.0test3bk3 report volanomark P N N 2.6.0test3 2.6.0test3bk3 report http://ltcperf.ncsa.uiuc.edu/data/2.6.0test3bk3/2.6.0test3vs2.6.0test3bk3/ Nightly Regression Summary for 2.6.0test3bk1 vs 2.6.0test3bk2 Benchmark Pass/Fail Improvements Regressions Results Results Summary        dbench.ext2 P N N 2.6.0test3bk1 2.6.0test3bk2 report dbench.ext3 P N Y 2.6.0test3bk1 2.6.0test3bk2 report dbench.jfs P N N 2.6.0test3bk1 2.6.0test3bk2 report dbench.reiser P N N 2.6.0test3bk1 2.6.0test3bk2 report dbench.xfs P N N 2.6.0test3bk1 2.6.0test3bk2 report kernbench P N Y 2.6.0test3bk1 2.6.0test3bk2 report lmbench P Y Y 2.6.0test3bk1 2.6.0test3bk2 report rawiobench P Y Y 2.6.0test3bk1 2.6.0test3bk2 report specjbb P Y Y 2.6.0test3bk1 2.6.0test3bk2 report specsdet P N N 2.6.0test3bk1 2.6.0test3bk2 report tbench P Y Y 2.6.0test3bk1 2.6.0test3bk2 report tiobench.ext2 P N N 2.6.0test3bk1 2.6.0test3bk2 report tiobench.ext3 P N Y 2.6.0test3bk1 2.6.0test3bk2 report tiobench.jfs P N N 2.6.0test3bk1 2.6.0test3bk2 report tiobench.reiser P N Y 2.6.0test3bk1 2.6.0test3bk2 report tiobench.xfs P N N 2.6.0test3bk1 2.6.0test3bk2 report volanomark P N N 2.6.0test3bk1 2.6.0test3bk2 report http://ltcperf.ncsa.uiuc.edu/data/2.6.0test3bk2/2.6.0test3bk1vs2.6.0test3bk2/ Nightly Regression Summary for 2.6.0test3 vs 2.6.0test3bk2 Benchmark Pass/Fail Improvements Regressions Results Results Summary        dbench.ext2 P N N 2.6.0test3 2.6.0test3bk2 report dbench.ext3 P N N 2.6.0test3 2.6.0test3bk2 report dbench.jfs P N N 2.6.0test3 2.6.0test3bk2 report dbench.reiser P N N 2.6.0test3 2.6.0test3bk2 report dbench.xfs P N N 2.6.0test3 2.6.0test3bk2 report kernbench P N N 2.6.0test3 2.6.0test3bk2 report lmbench P Y Y 2.6.0test3 2.6.0test3bk2 report rawiobench P Y Y 2.6.0test3 2.6.0test3bk2 report specjbb P Y Y 2.6.0test3 2.6.0test3bk2 report specsdet P N N 2.6.0test3 2.6.0test3bk2 report tbench P Y Y 2.6.0test3 2.6.0test3bk2 report tiobench.ext2 P N N 2.6.0test3 2.6.0test3bk2 report tiobench.ext3 P Y N 2.6.0test3 2.6.0test3bk2 report tiobench.jfs P N N 2.6.0test3 2.6.0test3bk2 report tiobench.reiser P N N 2.6.0test3 2.6.0test3bk2 report tiobench.xfs P N N 2.6.0test3 2.6.0test3bk2 report volanomark P N N 2.6.0test3 2.6.0test3bk2 report http://ltcperf.ncsa.uiuc.edu/data/2.6.0test3bk2/2.6.0test3vs2.6.0test3bk2/ Nightly Regression Summary for 2.6.0test3 vs 2.6.0test3bk1 Benchmark Pass/Fail Improvements Regressions Results Results Summary        dbench.ext2 P N N 2.6.0test3 2.6.0test3bk1 report dbench.ext3 P Y N 2.6.0test3 2.6.0test3bk1 report dbench.jfs P N N 2.6.0test3 2.6.0test3bk1 report dbench.reiser P N N 2.6.0test3 2.6.0test3bk1 report dbench.xfs P N N 2.6.0test3 2.6.0test3bk1 report kernbench P Y N 2.6.0test3 2.6.0test3bk1 report lmbench P Y Y 2.6.0test3 2.6.0test3bk1 report rawiobench P Y Y 2.6.0test3 2.6.0test3bk1 report specjbb P N Y 2.6.0test3 2.6.0test3bk1 report specsdet P N N 2.6.0test3 2.6.0test3bk1 report tbench P N Y 2.6.0test3 2.6.0test3bk1 report tiobench.ext2 P N N 2.6.0test3 2.6.0test3bk1 report tiobench.ext3 P Y N 2.6.0test3 2.6.0test3bk1 report tiobench.jfs P N N 2.6.0test3 2.6.0test3bk1 report tiobench.reiser P Y N 2.6.0test3 2.6.0test3bk1 report tiobench.xfs P Y N 2.6.0test3 2.6.0test3bk1 report volanomark P N N 2.6.0test3 2.6.0test3bk1 report http://ltcperf.ncsa.uiuc.edu/data/2.6.0test3bk1/2.6.0test3vs2.6.0test3bk1/ Nightly Regression Summary for 2.6.0test2 vs 2.6.0test2mm5 Benchmark Pass/Fail Improvements Regressions Results Results Summary        dbench.ext2 P N N 2.6.0test2 2.6.0test2mm5 report dbench.ext3 P Y N 2.6.0test2 2.6.0test2mm5 report dbench.jfs P N N 2.6.0test2 2.6.0test2mm5 report dbench.reiser P Y N 2.6.0test2 2.6.0test2mm5 report dbench.xfs P Y N 2.6.0test2 2.6.0test2mm5 report kernbench P N N 2.6.0test2 2.6.0test2mm5 report lmbench P Y Y 2.6.0test2 2.6.0test2mm5 report rawiobench P Y Y 2.6.0test2 2.6.0test2mm5 report specjbb P Y Y 2.6.0test2 2.6.0test2mm5 report specsdet P N Y 2.6.0test2 2.6.0test2mm5 report tbench P Y Y 2.6.0test2 2.6.0test2mm5 report tiobench.ext2 P N N 2.6.0test2 2.6.0test2mm5 report tiobench.ext3 P Y Y 2.6.0test2 2.6.0test2mm5 report tiobench.jfs P Y Y 2.6.0test2 2.6.0test2mm5 report tiobench.reiser P Y Y 2.6.0test2 2.6.0test2mm5 report tiobench.xfs P Y Y 2.6.0test2 2.6.0test2mm5 report volanomark P N Y 2.6.0test2 2.6.0test2mm5 report http://ltcperf.ncsa.uiuc.edu/data/2.6.0test2mm5/2.6.0test2vs2.6.0test2mm5/ Mark 
From: Steven Pratt <slpratt@au...>  20030822 20:30:49

Mark Peloquin wrote: > Been awhile since results where posted, therefore this is a little long. > > > Nightly Regression Summary for 2.6.0test3 vs 2.6.0test3mm3 > > Benchmark Pass/Fail Improvements Regressions > Results Results Summary >     >    > dbench.ext2 P N N 2.6.0test3 > 2.6.0test3mm3 report > dbench.ext3 P N Y 2.6.0test3 > 2.6.0test3mm3 report The ext3 dbench regression is very significant for multi threaded 193 > 118. Looks like this regression first showed up in mm1 and does not exist in any of the bk trees. http://ltcperf.ncsa.uiuc.edu/data/historygraphs/dbench.ext3.throughput.plot.16.png > > volanomark P N Y 2.6.0test3 > 2.6.0test3mm3 report Volanomark is significant as well. 10% drop in mm tree. This one also appeared to show up in mm1 although it was a 14% drop then so mm3 actually looks a little better. There were build errors on mm2 run so I don't have that data at this time. Following link illustrates the drop in mm tree for volanomark. http://ltcperf.ncsa.uiuc.edu/data/historygraphs/volanomark.throughput.plot.1.png SpecJBB2000 for high warehouses also took a bit hit. Probably the same root cause as volanomark. Here is the history plot for the 19 warehouse run. http://ltcperf.ncsa.uiuc.edu/data/historygraphs/specjbb.results.avg.plot.19.png Huge spike in idle time. http://ltcperf.ncsa.uiuc.edu/data/historygraphs/specjbb.utilization.idle.avg.plot.19.png > > http://ltcperf.ncsa.uiuc.edu/data/2.6.0test3mm3/2.6.0test3vs2.6.0test3mm3/ > Steve 
From: Andrew Morton <akpm@os...>  20030822 21:07:05

Gack, what happened to reiserfs? http://ltcperf.ncsa.uiuc.edu/data/historygraphs/dbench.reiser.throughput.plot.16.png Steven Pratt <slpratt@...> wrote: > > Mark Peloquin wrote: > > > Been awhile since results where posted, therefore this is a little long. > > > > > > Nightly Regression Summary for 2.6.0test3 vs 2.6.0test3mm3 > > > > Benchmark Pass/Fail Improvements Regressions > > Results Results Summary > >     > >    > > dbench.ext2 P N N 2.6.0test3 > > 2.6.0test3mm3 report > > dbench.ext3 P N Y 2.6.0test3 > > 2.6.0test3mm3 report > > The ext3 dbench regression is very significant for multi threaded 193 > > 118. Looks like this regression first showed up in mm1 and does not > exist in any of the bk trees. > > http://ltcperf.ncsa.uiuc.edu/data/historygraphs/dbench.ext3.throughput.plot.16.png > > > > > volanomark P N Y 2.6.0test3 > > 2.6.0test3mm3 report > > Volanomark is significant as well. 10% drop in mm tree. This one also > appeared to show up in mm1 although it was a 14% drop then so mm3 > actually looks a little better. There were build errors on mm2 run so I > don't have that data at this time. > Following link illustrates the drop in mm tree for volanomark. > > http://ltcperf.ncsa.uiuc.edu/data/historygraphs/volanomark.throughput.plot.1.png > > > SpecJBB2000 for high warehouses also took a bit hit. Probably the same > root cause as volanomark. > Here is the history plot for the 19 warehouse run. > > http://ltcperf.ncsa.uiuc.edu/data/historygraphs/specjbb.results.avg.plot.19.png > > Huge spike in idle time. > http://ltcperf.ncsa.uiuc.edu/data/historygraphs/specjbb.utilization.idle.avg.plot.19.png > > > > > http://ltcperf.ncsa.uiuc.edu/data/2.6.0test3mm3/2.6.0test3vs2.6.0test3mm3/ > > > > > Steve > 
From: Steven Pratt <slpratt@au...>  20030822 22:18:17

Andrew Morton wrote: >Gack, what happened to reiserfs? > >http://ltcperf.ncsa.uiuc.edu/data/historygraphs/dbench.reiser.throughput.plot.16.png > > There was a corresponding increase in reiser tio seq write throughput in the same kernel so I thought it might be intentional. I wrote this up and thought I had sent it to reiserfs list but never heard anything, but now I don't see my post in the archives, so maybe it never made it. So I'll cross post this over there to see what is up. http://ltcperf.ncsa.uiuc.edu/data/historygraphs/tiobench.reiser.Sequential_Write_thru.plot.64.png Steve 
From: Hans Reiser <reiser@na...>  20030823 02:18:27

Steven Pratt wrote: > Andrew Morton wrote: > >> Gack, what happened to reiserfs? >> >> http://ltcperf.ncsa.uiuc.edu/data/historygraphs/dbench.reiser.throughput.plot.16.png >> >> >> > There was a corresponding increase in reiser tio seq write throughput > in the same kernel so I thought it might be intentional. I wrote this > up and thought I had sent it to reiserfs list but never heard > anything, but now I don't see my post in the archives, so maybe it > never made it. So I'll cross post this over there to see what is up. > > http://ltcperf.ncsa.uiuc.edu/data/historygraphs/tiobench.reiser.Sequential_Write_thru.plot.64.png > > > Steve > > > > thanks for finding this. I'll say more when we have isolated the cause.  Hans 
From: Oleg Drokin <green@na...>  20030825 22:40:50

Hello! On Fri, Aug 22, 2003 at 05:17:44PM 0500, Steven Pratt wrote: > >Gack, what happened to reiserfs? > >http://ltcperf.ncsa.uiuc.edu/data/historygraphs/dbench.reiser.throughput.plot.16.png Yeah, that's weird. > There was a corresponding increase in reiser tio seq write throughput in > the same kernel so I thought it might be intentional. I wrote this up In fact, there were only two changes to reiserfs in between 2.5.75 and 2.6.0test1: ChangeSet@..., 20030711 00:01:3307:00, akpm@... [PATCH] reiserfs dirty memory accounting fix and ChangeSet@..., 20030711 00:01:4007:00, akpm@... [PATCH] fix reiserfs for 64bit arches None of which should even have any effect on dbench throughput. > and thought I had sent it to reiserfs list but never heard anything, but > now I don't see my post in the archives, so maybe it never made it. So > I'll cross post this over there to see what is up. > http://ltcperf.ncsa.uiuc.edu/data/historygraphs/tiobench.reiser.Sequential_Write_thru.plot.64.png Hm, I guess all these changes in throughput were because of some VM/VFS change, not because of reiserfs change. Now perhaps I will try to isolate what change was that. Bye, Oleg 
From: Andrew Morton <akpm@os...>  20030822 21:02:28

Steve, thanks for summarizing the results in this manner. Frankly I wouldn't have noticed *any* of these regressions had you not performed this analysis  there's a lot of valueadd on top of the raw numbers here. I'd be suspecting the CPU scheduler changes for most things, but perhaps not the dbenchonext3 problem. RAMonly dbench shouldn't regress like that; I can take a look at it. volanomark and specjbb are harder for me to analyse. Could someone at your end please do a quick A/B test? Grab a current Linus tree (2.6.0test3/bk8 will do) and test volanomark or specjbb with and without http://www.zip.com.au/~akpm/linux/patches/schedrollup.patch I wouldn't do an exhaustive test: just enough to be able to say "yup, that's the culprit" or "no, it's something else". Thanks. Steven Pratt <slpratt@...> wrote: > > Mark Peloquin wrote: > > > Been awhile since results where posted, therefore this is a little long. > > > > > > Nightly Regression Summary for 2.6.0test3 vs 2.6.0test3mm3 > > > > Benchmark Pass/Fail Improvements Regressions > > Results Results Summary > >     > >    > > dbench.ext2 P N N 2.6.0test3 > > 2.6.0test3mm3 report > > dbench.ext3 P N Y 2.6.0test3 > > 2.6.0test3mm3 report > > The ext3 dbench regression is very significant for multi threaded 193 > > 118. Looks like this regression first showed up in mm1 and does not > exist in any of the bk trees. > > http://ltcperf.ncsa.uiuc.edu/data/historygraphs/dbench.ext3.throughput.plot.16.png > > > > > volanomark P N Y 2.6.0test3 > > 2.6.0test3mm3 report > > Volanomark is significant as well. 10% drop in mm tree. This one also > appeared to show up in mm1 although it was a 14% drop then so mm3 > actually looks a little better. There were build errors on mm2 run so I > don't have that data at this time. > Following link illustrates the drop in mm tree for volanomark. > > http://ltcperf.ncsa.uiuc.edu/data/historygraphs/volanomark.throughput.plot.1.png > > > SpecJBB2000 for high warehouses also took a bit hit. Probably the same > root cause as volanomark. > Here is the history plot for the 19 warehouse run. > > http://ltcperf.ncsa.uiuc.edu/data/historygraphs/specjbb.results.avg.plot.19.png > > Huge spike in idle time. > http://ltcperf.ncsa.uiuc.edu/data/historygraphs/specjbb.utilization.idle.avg.plot.19.png > > > > > http://ltcperf.ncsa.uiuc.edu/data/2.6.0test3mm3/2.6.0test3vs2.6.0test3mm3/ > > > > > Steve > 
From: Steven Pratt <slpratt@au...>  20030822 22:08:57

Andrew Morton wrote: >Steve, > >thanks for summarizing the results in this manner. Frankly I wouldn't have >noticed *any* of these regressions had you not performed this analysis  >there's a lot of valueadd on top of the raw numbers here. > >I'd be suspecting the CPU scheduler changes for most things, but perhaps >not the dbenchonext3 problem. > >RAMonly dbench shouldn't regress like that; I can take a look at it. > >volanomark and specjbb are harder for me to analyse. Could someone at your >end please do a quick A/B test? > >Grab a current Linus tree (2.6.0test3/bk8 will do) and test volanomark or >specjbb with and without > > http://www.zip.com.au/~akpm/linux/patches/schedrollup.patch > >I wouldn't do an exhaustive test: just enough to be able to say "yup, >that's the culprit" or "no, it's something else". > > First glance says the *is* the problem. Volanomark results: test3bk8 without patch:41059 40791 40975 41723 41719 AVG: 41253 STDDEV: 0% test3bk8 with patch :37195 36640 37615 37341 36691 AVG: 37096 STDDEV: 1% test3mm3 results :36768 36480 36788 36413 35974 AVG: 36484 STDDEV: 0% I have jbb running, but it won't finish before I leave tonight. Let me know if there is anything else I can do to help out. Steve 
From: Steven Pratt <slpratt@au...>  20030825 23:10:42

Andrew Morton wrote: >Steve, > >thanks for summarizing the results in this manner. Frankly I wouldn't have >noticed *any* of these regressions had you not performed this analysis  >there's a lot of valueadd on top of the raw numbers here. > >I'd be suspecting the CPU scheduler changes for most things, but perhaps >not the dbenchonext3 problem. > >RAMonly dbench shouldn't regress like that; I can take a look at it. > >volanomark and specjbb are harder for me to analyse. Could someone at your >end please do a quick A/B test? > >Grab a current Linus tree (2.6.0test3/bk8 will do) and test volanomark or >specjbb with and without > > http://www.zip.com.au/~akpm/linux/patches/schedrollup.patch > >I wouldn't do an exhaustive test: just enough to be able to say "yup, >that's the culprit" or "no, it's something else". > > Just a quick followup. Specjbb did show similar degredation with the application of the schedrollup patch on test3bk8. Steve 