x/perf/benchstat: bogus "no statistical difference" report when times are the same #19634
Comments
|
Also see Alberto's suggested fix at rsc/benchstat#7 (comment). |
|
@mvdan I never sent a patch because I'm not 100% sure the explanation I gave on the old repo's issues tracker is correct; anyway if you are convinced it is correct you can send a change, if you'd like to. |
|
Ah, best to not send it then. |
|
/cc @aclements |
|
The problem here isn't the significance test, it's the outlier rejection combined with the small sample size. As an order test, Mann-Whitney has a floor on the p value that depends on the number of samples (not their values). p=0.079 is simply the lowest p-value you can get with n=4, m=5. The significance test isn't failing. It's genuinely saying that with so few samples, the chance of getting that order randomly is 0.079. If you change the 114 ns/op to 115 ns/op, there's even less variance, but now outlier rejection doesn't kick in, so you get n=5, m=5 and a p-value of 0.008, which is considered significant. I think the real bug here is that we're doing outlier rejection before computing an order statistic. We probably shouldn't do that. But if we still want to do outlier rejection for computing the mean ± x%, then I'm not sure how to present the sample size. Maybe we shouldn't be doing outlier rejection for that either. Perhaps we should be reporting a trimmed mean and its standard error? Independently, perhaps benchstat should report when the sample sizes are too small to ever get a significant result. |
|
I hit the same or similar problem with "macro"-benchmark. My data: old.txt: new.txt:
|
|
@AlekSi that is expected - you'll need to run each benchmark multiple times - that's what the |
|
Thank you. I wish it were more clear in the command output and documentation, though. |
|
Change https://golang.org/cl/309969 mentions this issue: |
This is a complete rewrite of benchstat. Basic usage remains the same, as does the core idea of showing statistical benchmark summaries and A/B comparisons in a table, but there are several major improvements. The statistics is now more robust. Previously, benchstat used IQR-based outlier rejection, showed the mean of the reduced sample, its range, and did a non-parametric difference-of-distribution test on reduced samples. Any form of outlier rejection must start with distributional assumptions, in this case assuming normality, which is generally not sound for benchmark data. Hence, now benchstat does not do any outlier rejection. As a result, it must use robust summary statistics as well, so benchstat now uses the median and confidence interval of the median as summary statistics. Benchstat continues to use the same Mann-Whitney U-test for the delta, but now does it on the full samples since the U-test is already non-parametric, hence increasing the power of this test. As part of these statistical improvements, benchstat now detects and warns about several common mistakes, such as having too few samples for meaningful statistical results, or having incomparable geomeans. The output format is more consistent. Previously, benchstat transformed units like "ns/op" into a metric like "time/op", which it used as a column header; and a numerator like "sec", which it used to label each measurement. This was easy enough for the standard units used by the testing framework, but was basically impossible to generalize to custom units. Now, benchstat does unit scaling, but otherwise leaves units alone. The full (scaled) unit is used as a column header and each measurement is simply a scaled value shown with an SI prefix. This also means that the text and CSV formats can be much more similar while still allowing the CSV format to be usefully machine-readable. Benchstat will also now do A/B comparisons even if there are more than two inputs. It shows a comparison to the base in the second and all subsequent columns. This approach is consistent for any number of inputs. Benchstat now supports the full Go benchmark format, including sophisticated control over exactly how it structures the results into rows, columns, and tables. This makes it easy to do meaningful comparisons across benchmark data that isn't simply structured into two input files, and gives significantly more control over how results are sorted. The default behavior is still to turn each input file into a column and each benchmark into a row. Fixes golang/go#19565 by showing all results, even if the benchmark sets don't match across columns, and warning when geomean sets are incompatible. Fixes golang/go#19634 by no longer doing outlier rejection and clearly reporting when there are not enough samples to do a meaningful difference test. Updates golang/go#23471 by providing more through command documentation. I'm not sure it quite fixes this issue, but it's much better than it was. Fixes golang/go#30368 because benchstat now supports filter expressions, which can also filter down units. Fixes golang/go#33169 because benchstat now always shows file configuration labels. Updates golang/go#43744 by integrating unit metadata to control statistical assumptions into the main tool that implements those assumptions. Fixes golang/go#48380 by introducing a way to override labels from the command line rather than always using file names. Change-Id: Ie2c5a12024e84b4918e483df2223eb1f10413a4f
Moving from rsc/benchstat#7, which now appears to be a dead issue tracker.
In short - if you happen to get benchmark time numbers that happen to be the same,
benchstatseems to discard them.old.txt:
new.txt (note that all the times are the same: 78.8 ns/op):
benchstat old.txt new.txt gives:
i.e. reports "no statistically significant improvement", which is clearly wrong.
cc @ALTree @rsc
The text was updated successfully, but these errors were encountered: