These measurements were taken using the schedstats patch, combined with a graphing tool still being refined (to be made available soon). In the test5* cases, samples were taken every 30 seconds, and on some of the shorter benchmarks that provides only a smattering of data points. In test6* cases, samples were taken every 5 seconds but then became prone to "observational error"; that is, the process that took the stats tended to get them a little earlier or a little later than five seconds as it itself was subject to scheduling irregularities. This was corrected in the test6* samples by combining six 5-second samples into a single 30 second sample, which tended to smooth out the irregularities. A future schedstat will incorporate a way to utilize finer granularity without being subject to scheduling skew.