On Fri, Aug 24, 2018 at 12:45:10PM -0400, Derrick Stolee wrote: > Here are the numbers for Linux: > > | | v2.18.0 | v2.19.0-rc0 | HEAD | > |------|----------|-------------|--------| > | | 86.5 | 70.739 | 57.266 | > | | 60.582 | 101.928 | 56.641 | > | | 58.964 | 60.139 | 60.258 | > | | 59.47 | 61.141 | 58.213 | > | | 62.554 | 60.73 | 84.54 | > | | 59.139 | 85.424 | 57.745 | > | | 58.487 | 59.31 | 59.979 | > | | 58.653 | 69.845 | 60.181 | > | | 58.085 | 102.777 | 61.455 | > | | 58.304 | 60.459 | 62.551 | > | Max | 86.5 | 102.777 | 84.54 | > | Min | 58.085 | 59.31 | 56.641 | > | Avg* | 59.51913 | 71.30063 | 59.706 | > | Med | 59.0515 | 65.493 | 60.08 | Thanks. The median ones are the most interesting, I think (and show a very satisfying recovery from my patch). I'm surprised at the variance, especially in your v2.19 runs. It makes me distrust the mean (and implies to me we could do better by throwing away outliers based on value, not just the single high/low; or just using the median also solves that). The variance in the v2.18 column is what I'd expect based on past experience (slow cold cache to start, then a few percent change run-to-run). -Peff
