On Thu, 26 Jan 2023 11:25:29 GMT, Claes Redestad <[email protected]> wrote:
> As a rule of thumb if you get the approximate same results when running the > same benchmark back to back and the difference between your baseline and test > results is decidedly larger than the sum of the errors then you can have a > reasonably high confidence in the results. > > If errors are too high for comfort (or effect too small) you should first > consider if your benchmarking system can be made quieter (e.g. turning off > browsers, slack, IDEs, turn off hyperthreading, turbo boost, apply firmware > updates, put the box in a freezer, don't move, don't make a sound...!), then > if you're still not getting comfortably consistent results you can increase > the amount of forks, warmup and measurement iterations you're running with. So I did some runs yesterday and then a few more today and it seems like on my machine the average time dropped around 15% for the `EncodeDecode` test, It seemed pretty consistent with that though I'm sure I was far from thorough ------------- PR: https://git.openjdk.org/jdk/pull/12122
