If you are looking at shaving sub 20ms numbers, be aware you may need to pull tricks on macos especially to get real numbers.
I've run into some memory corruption with it sometimes, though, so be wary of that. Emerge tools has an alternative for iOS at least, maybe one day they'll port it to mac.
For my purposes I use it all the time though, quick and easy sanity-check.
$ hyperfine --warmup 3 './hello-world-bin-sh.sh' './hello-world-env-python3.py'
Benchmark 1: ./hello-world-bin-sh.sh
Time (mean ± σ): 1.3 ms ± 0.4 ms [User: 1.0 ms, System: 0.5 ms]
...
Benchmark 2: ./hello-world-env-python3.py
Time (mean ± σ): 43.1 ms ± 1.4 ms [User: 33.6 ms, System: 8.4 ms]
...---
https://abuisman.com/posts/developer-tools/quick-page-benchm...
As mentioned here in the thread, when you want to go into the single ms optimisations it is not the best approach since there is a lot of overhead especially the way I demonstrate here, but it works very well for some sanity checks.
But for everything in the right range (milliseconds, seconds, minutes or above), hyperfine is well suited.
If you see any reason for putting “statistical” in quotes, please let us know. hyperfine does not run a lot of tests, but it does try to find outliers in your measurements. This is really valuable in some cases. For example: we can detect when the first run of your program takes much longer than the rest of the runs. We can then show you a warning to let you know that you probably want to either use some warmup runs, or a "--prepare" command to clean (OS) caches if you want a cold-cache benchmark.
> But there’s no good way to say “just run it for 5 seconds and give me the best answer you can”.
What is the "best answer you can"?
> It’s very much designed for nanosecond to low microsecond benchmarks.
Absolutely not. With hyperfine, you can not measure execution times in the "low microsecond" range, let alone nanosecond range. See also my other comment.
hyperfine -N -- ls "$dir" \; my_ls "$dir"Looks fine to me. Obviously it's too late to undo that mistake, but a new flag to enable new behavior wouldn't hurt anyone.
Is it, though?
What I would expect a system like this to have, at a minimum:
* Robust statistics with p-values (not just min/max, compensation for multiple hypotheses, no Gaussian assumptions)
* Multiple stopping points depending on said statistics.
* Automatic isolation to the greatest extent possible (given appropriate permissions)
* Interleaved execution, in case something external changes mid-way.
I've tried to understand what the issue is (played with resigning executables etc) but it's literally something about the inode of the executable itself. Most likely part of the OSX security system.
Back at the time I needed it, it had peak memory usage - hyperfine was not able to show it. Maybe this had changed by now.
Back in the day my goal for Advent of Code was to run all solutions in under 1 second total. Hyperfine would take like 30 minutes to benchmark a 1 second runtime.
It was hyper frustrating. I could not find a good way to get Hyperfine to do what I wanted.
> I could not find a good way to get Hyperfine to do what I wanted
This is all documented here: https://github.com/sharkdp/hyperfine/tree/master?tab=readme-... under "Basic benchmarks". The options to control the amount of runs are also listed in `hyperfine --help` and in the man page. Please let us know if you think we can improve the documentation / discovery of those options.
This is not included in the core of hyperfine, but we do have scripts to compute "advanced" statistics, and to perform t-tests here: https://github.com/sharkdp/hyperfine/tree/master/scripts
Please feel free to comment here if you think it should be included in hyperfine itself: https://github.com/sharkdp/hyperfine/issues/523
> Automatic isolation to the greatest extent possible (given appropriate permissions)
This sounds interesting. Please feel free to open a ticket if you have any ideas.
> Interleaved execution, in case something external changes mid-way.
Please see the discussion here: https://github.com/sharkdp/hyperfine/issues/21
> It just… runs things N times and then does a naïve average/min/max?
While there is nothing wrong with computing average/min/max, this is not all hyperfine does. We also compute modified Z-scores to detect outliers. We use that to issue warnings, if we think the mean value is influenced by them. We also warn if the first run of a command took significantly longer than the rest of the runs and suggest counter-measures.
Depending on the benchmark I do, I tend to look at either the `min` or the `mean`. If I need something more fine-grained, I export the results and use the scripts referenced above.
> At that rate, one could just as well use a shell script and eyeball the results.
Statistical analysis (which you can consider to be basic) is just one reason why I wrote hyperfine. The other reason is that I wanted to make benchmarking easy to use. I use warmup runs, preparation commands and parametrized benchmarks all the time. I also frequently use the Markdown export or the JSON export to generate graphs or histograms. This is my personal experience. If you are not interested in all of these features, you can obviously "just as well use a shell script".
t-tests run afoul of the “no Gaussian assumptions”, though. Distributions arising from benchmarking frequently has various forms of skew which messes up t-tests and gives artificially narrow confidence intervals.
(I'll gladly give you credit for your outlier detection, though!)
>> Automatic isolation to the greatest extent possible (given appropriate permissions) > This sounds interesting. Please feel free to open a ticket if you have any ideas.
Off the top of my head, some option that would:
* Bind to isolated CPUs, if booted with it (isolcpus=) * Binding to a consistent set of cores/hyperthreads (the scheduler frequently sabotages benchmarking, especially if your cores are have very different maximum frequency) * Warns if thermal throttling is detected during the run * Warns if an inappropriate CPU governor is enabled * Locks the program into RAM (probably hard to do without some sort of help from the program) * Enables realtime priority if available (e.g., if isolcpus= is not enabled, or you're not on Linux)
Of course, sometimes you would _want_ to benchmark some of these effects, and that's fine. But most people probably won't, and won't know that they exist. I may easily have forgotten some.
On the flip side (making things more random as opposed to less), something that randomizes the initial stack pointer would be nice, as I've sometimes seen this go really, really wrong (renaming a binary from foo to foo_new made it run >1% slower!).
This is something we do already. We set a `HYPERFINE_RANDOMIZED_ENVIRONMENT_OFFSET` environment variable with a random-length value: https://github.com/sharkdp/hyperfine/blob/87d77c861f1b6c761a...
Current defaults: "By default, it will perform at least 10 benchmarking runs and measure for at least 3 seconds." If your program takes 1s to run, it should take 10 seconds to benchmark.
Is it possible that your program was waiting for input that never came? One "gotcha" is that it expects each argument to be a full program, so if you ran `hyperfine ./a.out input.txt`, it will first bench a.out with no args, then try to bench input.txt (which will fail). If a.out reads from stdin when no argument is given, then it would hang forever, and I can see why you'd give up after a half hour.
Pointing this out because I myself appreciate comments that do this.
For myself, `fd` is the one most incorporated into my own "toolbox" -- used it this morning prior to seeing this thread on hyperfine! So, thanks for all that, sharkdp if you're reading!
Ok, end OT-ness.
Even with that I reach for `fd` for some of its quality-of-life features: respecting .gitignore, its speed, regex-ability. (Though not its choices with color; I am a pretty staunch "--color never" person, for better or worse!)
Anyway, that actually points to another good thing about sharkdp's tools: they have good man pages!!
> I stand firm in my belief that unless you can prove how CLT applies to your input distributions, you should not assume normality. And if you don't know what you are doing, stop reporting means.
I agree. My research group stopped using Hyperfine because it ranks benchmarked commands by mean, and provides standard deviation as a substitute for a confidence measure. These are not appropriate for heavy-tailed, skewed, and otherwise non-normal distributions.
It's easy to demonstrate that most empirical runtime distributions are not normal. I wrote BestGuess [0] because we needed a better benchmarking tool. Its analysis provides measures of skew, kurtosis, and Anderson-Darling distance from normal, so that you can see how normal or not is your distribution. It ranks benchmark results using non-parametric methods. And, unlike many tools, it saves all of the raw data, making it easy to re-analyze later.
My team also discovered that Hyperfine's measurements are a bit off. It reports longer run times than other tools, including BestGuess. I believe this is due to the approach, which is to call getrusage(), then fork/exec the program to be measured, then call getrusage() again. The difference in user and system times is reported as the time used by the benchmarked command, but unfortunately this time also includes cycles spent in the Rust code for managing processes (after the fork but before the exec).
BestGuess avoids external libraries (we can see all the relevant code), does almost nothing after the fork, and uses wait4() to get measurements. The one call to wait4() gives us what the OS measured by its own accounting for the benchmarked command.
While BestGuess is still a work in progress (not yet at version 1.0), my team has started using it regularly. I plan to continue its development, and I'll write it up soon at [1].
[0] https://gitlab.com/JamieTheRiveter/bestguess [1] https://jamiejennings.com