Tag: benchmarking

Slides: Stop Guessing and Start Measuring (Poly-Version)

Slides: Stop Guessing and Start Measuring (Poly-Version)

Hello from the amazing Polyconf! I just gave my Stop Guessing and Start Measuring talk and if you are thinking “why do you post the slides of this SO MANY TIMES”, well the first one was an Elixir version, then a Ruby + Elixir version and now we are at a Poly version. The slides are mostly different and I’d say about ~50% of them are new. New topics covered include:

  • MJIT – what’s wrong with the benchmarks – versus TruffleRuby
  • JavaScript!
  • other nice adjustments

The all important video isn’t in the PDF export but you can see a big part of it on Instagram.

You can view the slides here or on speakerdeck, slideshare or PDF.


“What’s the fastest way of doing this?” – you might ask yourself during development. Sure, you can guess, your intuition might be correct – but how do you know? Benchmarking is here to give you the answers, but there are many pitfalls in setting up a good benchmark and analyzing the results. This talk will guide you through, introduce best practices, and surprise you with some unexpected benchmarking results. You didn’t think that the order of arguments could influence its performance…or did you?



Slides: How fast is it really? Benchmarking in Elixir

I’m at Elixirlive in Warsaw right now and just gave a talk. This talk is about benchmarking – the greater concepts but concrete examples are in Elixir and it works with my very own library benchee to also show some surprising Elixir benchmarks. The concepts are applicable in general and it also gets into categorizing benchmarks into micro/macro/application etc.

If you’ve been here and have feedback – positive or negative. Please tell me 🙂

Slides are available as PDF, speakerdeck and slideshare.


“What’s the fastest way of doing this?” – you might ask yourself during development. Sure, you can guess what’s fastest or how long something will take, but do you know? How long does it take to sort a list of 1 Million elements? Are tail-recursive functions always the fastest?

Benchmarking is here to answer these questions. However, there are many pitfalls around setting up a good benchmark and interpreting the results. This talk will guide you through, introduce best practices and show you some surprising benchmarking results along the way.

Released: benchee 0.6.0, benchee_csv 0.5.0, benchee_json and benchee_html – HTML reports and nice graphs!

Released: benchee 0.6.0, benchee_csv 0.5.0, benchee_json and benchee_html – HTML reports and nice graphs!

The last days I’ve been hard at work to polish up and finish releases of benchee (0.6.0 – Changelog), benchee_csv (0.5.0 – Changelog) as well as the initial releases of benchee_html and benchee_json!

I’m the proudest and happiest of finally getting benchee_html out of the door along with great HTML reports including plenty of graphs and the ability to export them! You can check out the example online report or glance at this screenshot of it:

reportWhile benchee_csv had some mere updates for compatibility and benchee_json just transforms the general suite to JSON (which is then used in the HTML formatter) I’m particularly excited about the big new features in benchee and of course benchee_html!


The 0.6.0 is probably the biggest release of the “core” benchee library with some needed API changes and great features.

New run API – options last as keyword list

The “old” way you’d optionally pass in options as the first argument into run as a map and then define the jobs to benchmark in another map. I did this because in my mind the configuration comes first and maps are much easier to work with through pattern matching as opposed to keyword lists. However, having an optional first argument already felt kind of weird…

Thing is, that’s not the most elixir way to do this. It is rather conventional to pass in options as the last argument and as a keyword list. After voicing my concerns in the elixirforum, the solution was to allow passing in options as keyword lists but convert to maps internally to still have the advantage of good pattern matching among other advantages.

The old style still works (thanks to pattern matching!) – but it might get deprecated in the future. In this process though the run interface of the very first version of run, which used a list of tuples, doesn’t work anymore 😦

Multiple inputs

The great new feature is that benchee now supports multiple inputs – so that in one suite you can run the same functions against multiple different inputs. That is important as functions can behave very differently on inputs of different sizes or a different structure. Therefore it’s good to check the functions against multiple inputs. The feature was inspired by a discussion on an elixir issue with José Valim.

So what does this look like? Here it goes:

The hard thing about it was that it changed how benchmarking results had to be represented internally, as another level to represent the different inputs was needed. This lead to quite some work both in benchee and in plugins – but in the end it was all worth it 🙂


This has been in the making for way too long, should have released a month or 2 ago. But now it’s here! It provides a nice HTML table and four different graphs – 2 for comparing the different benchmarking jobs and 2 graphs for each individual job to take a closer look at the distribution of run times of this particular job. There is a wiki page at benchee_html to discern between the different graphs highlighting what they might be useful for. You can also export PNG images of the graphs at click of a simple icon 🙂

Wonder how to use it? Well it was already shown earlier in this post when showing off the new API. You just specify the formatters and the file where it should be written to 🙂

But without further ado you can check out the sample report or just take a look at these images 🙂


raw_run_timesClosing Thoughts

Hope you enjoy benchmarking, with different inputs and then see great reports of them. Let me know what you like about benchee or what you don’t like about it and what could be better.

Benchee 0.4.0 released – adjust what is printed

Today I made a little 0.4.0 release of my elixir benchmarking library benchee. As always the Changelog has all the details.

This release mainly focusses on making all non essential output that benchee produces optional. This is mostly rooted in user feedback of people who wanted to disable the fast execution warnings or the comparison report. I decided to go full circle and also make it configurable if benchee prints out which job it is currently benchmarking or if the general configuration information is printed. I like this sort of verbose information and progress feedback – but clearly it’s not to everyone’s taste and that’s just fine 🙂

So what’s next for benchee? As a keen github observer might have noticed I’ve taken a few stabs at rendering charts in HTML + JS for benchee and in the process created benchee_json. I’m a bit dissatisfied as of now, as I’d really want to have graphs showing error bars and that seems to be harder to come by than I thought. After D3 and chart.js I’ll probably give highcharts a stab now. However, just reading the non-commercial terms again I’m not too sure if it’s good in all sense (e.g. what happens if someone in a commercial corporation uses and generates the HTML?). Oh, but the wonders of the Internet in a new search I found plotly which seems to have some great error bars support.

Other future plans include benchmarking with multiple input sizes to see how different approaches perform or the good old topic of lessening the impact of garbage collection 🙂


Benchee 0.3.0 released – formatters, parallel benchmarking & more

Yesterday I released benchee 0.3.0! Benchee is a tool for (micro) benchmarking in elixir focussing on being simple, extensible and to provide you with good statistics. You can refer to the Changelog for detailed information about the changes. This post will look at the bigger changes and also give a bit of the why for the new features and changes.

Multiple formatters

Arguably the biggest feature in Benchee 0.3.0 is that it is now easy and built-in to configure multiple formatters for a benchmarking suite. This means that first the benchmark is run, and then multiple formatters are run on the benchmarking results. This way you can get both the console output and the corresponding csv file using BencheeCSV. This was a pain point for me before, as you could either get one or the other or you needed to use the more verbose API.

You can also see the new output/1 methods at work, as opposed to format/1 they also really do the output themselves. BencheeCSV uses a custom configuration options to know which file to write to. This is also new, as now formatters have access to the full benchmarking suite, including configuration, raw run times and function definitions. This way they can be configured using configuration options they define themselves, or a plugin could graph all run times if it wanted to.

Of course, formatters default to just the built-in console formatter.

Parallel benchmarking

Another big addition is parallel benchmarking. In Elixir, this just feels natural to have. You can specify a parallel key in the configuration and that tells Benchee how many tasks should execute any given benchmarking job in parallel.

Of course, if you want to see how a system behaves under load – overloading might be exactly what you want to stress test the system. And this was exactly the reason why Leon contributed this change back to Benchee:

I needed to benchmark integration tests for a telephony system we wrote – with this system the tests actually interfere with each other (they’re using an Ecto repo) and I wanted to see how far I could push the system as a whole. Making this small change to Benchee worked perfectly for what I needed 🙂

(Of course it makes me extremely happy that people found adjusting Benchee for their use case simple, that’s one of the main goals of Benchee. Even better that it was contributed back ❤ )

If you want to see more information and detail about “to benchmark in parallel or not” you can check the Benchee wiki. Spoiler alert: The more parallel benchmarks run, the slower they get to an acceptable degree until the system is overloaded (more tasks execute in parallel than there are CPU cores to take care of them). Also deviation skyrockets.

While the effect seems not to be very significant for parallel: 2 on my system, the default in Benchee remains parallel: 1 for the mentioned reasons.

Print configuration information

Partly also due to the parallel change, Benchee wil now print a brief summary of the benchmarking suite before executing it.

tobi@happy ~/github/benchee $ mix run samples/run_parallel.exs

Benchmark suite executing with the following configuration:
warmup: 2.0s
time: 3.0s
parallel: 2
Estimated total run time: 10.0s

Benchmarking flat_map...
Benchmarking map.flatten...

Name                  ips        average    deviation         median
map.flatten       1268.15       788.55μs    (±13.94%)       759.00μs
flat_map           706.35      1415.72μs     (±8.56%)      1419.00μs

map.flatten       1268.15
flat_map           706.35 - 1.80x slower

This was done so that when people share their benchmarks online one can easily see the configuration they ran it with. E.g. was there any warmup time? Was the amount of parallel tasks too high and therefore the results are that bad?

It also prints an estimated total run time (number of jobs * (warmup + time)), so you know if there’s enough time to go and get a coffee before a benchmark finishes.

Map instead of a list of tuples

What is also marked as a “breaking” change in the Changelog is actually not THAT breaking. The main data structure handed to Benchee.run was changed to a map instead of a list of tuples and all corresponding data structures changed as well (important for plugins to know).

It used to be a list of tuples because of the possibility that benchmarks with the same name would override each other. However, having benchmarks with the same name is nonsensical as you can’t discern their results in the output any way. So, this now feels like a much more fitting data structure.

The old main data structure of a list of tuples still works and while I might remove it, I don’t expect me to right now as all that is required to maintain it is 4 lines of code. This makes duplicated names no longer working the only real deprecation, although one might even call it a feature 😉

Last, but not least, this release is the first one that got some community contributions in, which makes me extremely happy. So, thanks Alvin and Leon! 😀

Benchee 0.2.0 – warmup & nicer console output

Less than a week after the initial release of my benchmarking library Benchee there is a new version – 0.2.0! The details are in the Changelog. That’s the what, but what about the why?


Arguably the biggest change is introduction of a warmup phase to the benchmarks. That is the benchmark jobs are first run for some time without taking measurements to simulate a “warm” already running system. I didn’t think it’d be that important as the BEAM VM isn’t JITed (as opposed to the JVM) for all hat I know. It is important once benchmarks get to be “macro” – for instance databases usually respond faster once they got used to some queries and our webservers serve most of their time “hot”.

However, even in my micro benchmarks I noticed that it could have an effect when a benchmark was moved around (being run first versus being run last). So I don’t know to what effect, but at least to a small effect there is warmup now. If you don’t want warmup – just set warmup: 0.

Nicer console output

Name                                    ips        average    deviation         median
bodyrecusrive map                  40047.87        24.97μs    (±32.55%)        25.00μs
stdlib map                         39724.07        25.17μs    (±61.41%)        25.00μs
map tco no reverse                 36388.50        27.48μs    (±23.22%)        27.00μs
map with TCO and reverse           33309.43        30.02μs    (±45.39%)        29.00μs
map with TCO and ++                  465.25      2149.40μs     (±4.84%)      2138.00μs

bodyrecusrive map                  40047.87
stdlib map                         39724.07 - 1.01x slower
map tco no reverse                 36388.50 - 1.10x slower
map with TCO and reverse           33309.43 - 1.20x slower
map with TCO and ++                  465.25 - 86.08x slower

The ouput of numbers is now aligned right, which makes them easier to read and compare, as you can see orders of magnitude differences much more easily. Also the ugly empty line at the end of the output has been removed 🙂


This is the API incompatible change. It felt weird to me in version 0.1.0 that Benchee.benchmark would already run the function given to it. Now the jobs are defined through Benchee.benchmark and kept in a datastructure (similar to the one Benchee.run uses). Benchee.measure then runs the jobs and measures the outcome and provides them under the new run_times key instead of overriding the jobs key. This feels much nicer overall, of course the high level Benchee.run is unaffected by this.

These additions already nicely improve what Benchee can do and got a couple of items off my “I want to do this in benchee” bucket list. There’s still more to come 🙂

Introducing Benchee: simple and extensible benchmarking for Elixir

If you look around this blog it becomes pretty clear that I really love (micro) benchmarking. Naturally, while working more and more with Elixir (and loving it!) I wanted to benchmark something. Sadly, the existing options I found didn’t quite satisfy me. Be it for a different focus, missing statistics, lacking documentation or other things. So I decided to roll my own, it’s not like it’d be the first time.

Of course I tried extending existing solutions but very long functions, very scarce test coverage, lots of dead and outcommented code and a rotting PR later I decided it was time to create something new. So without further ado, please meet Benchee (of course available on hex)!

What’s great about Benchee?

Benchee is easy to use, well documented and can be extended (more on that in the following paragraphs). Benchee will run each benchmarking function you give it for a given amount of time and then compute statistics from it. Statistics is where it shines in my opinion. Benchee provides you with:

  • average run time (ok – yawn)
  • iterations per second, which is great for graphs etc. as higher is better here (as opposed to average run time)
  • standard deviation, an important value in my opinion as it gives you a feeling for how certain you can be about your measurements and how much they vary. Sadly, none of the elixir benchmarking tools I looked at supplied this value.
  • median, it’s basically the middle value of your distribution and is often cited as a value that reflects the “common” outcome better than average as it cuts out outliers. I never used a (micro) benchmarking tool that provided this value, but was often asked to provide it in my benchmarks. So here it is!

Also it gives a rather nice output on the console with headers so you know what is what. An example is further down but for now let’s talk design…

Designing a Benchmarking library

The design is influenced by my favourite ruby benchmarking library: benchmark-ips. Of course I wanted it to be more of an elixirish spin and offer more options.

A lot of elixir solutions used macros. I wanted something that works purely with functions, no tricks. When I started to learn more about functional programming one of the things that stuck with me the most was that functional programming is about a series of transformations. So what do these transformations look like for benchmarking?

  1. Create a basic benchmarking configuration with things like how long should the benchmark run, should GC be enabled etc.
  2. Run individual benchmarks and record their raw execution times
  3. Compute statistics based on these raw run times per benchmark
  4. Format the statistics to be suitable for output
  5. Put out the formatted statistics to the console, a file or whatever

So what do you now, that’s exactly what the API of Benchee looks like!

list = Enum.to_list(1..10_000)
map_fun = fn(i) -> [i, i * i] end

Benchee.init(%{time: 3})
|> Benchee.benchmark("flat_map", fn -> Enum.flat_map(list, map_fun) end)
|> Benchee.benchmark("map.flatten",
                     fn -> list |> Enum.map(map_fun) |> List.flatten end)
|> Benchee.statistics
|> Benchee.Formatters.Console.format
|> IO.puts

What’s great about this? Well it’s super flexible and flows nicely with the beloved elixir pipe operator.

Why is this flexible and extensible? Well, don’t like how Benchee runs the benchmarks? Sub in your own benchmarking function! Want more/different statistics? Go use your own function and compute your own! Want results to be displayed in a different format? Roll you own formatter! Or you just want to write the results to a file? Well, go ahead!

This is more than just cosmetics. It’d be easy to write a plugin that converts the results to some JSON format and then post them to a web service to gather benchmarking results or let it generate fancy graphs for you.

Of course, not everybody needs that flexibility. Some people might be scared away by the verboseness above. So there’s also a higher level interface that uses all the options you see above and condenses them down to one function call to efficiently define your benchmarks:

list = Enum.to_list(1..10_000)
map_fun = fn(i) -> [i, i * i] end

Benchee.run(%{time: 3},
             [{"flat_map", fn -> Enum.flat_map(list, map_fun) end},
              fn -> list |> Enum.map(map_fun) |> List.flatten end}])

Let’s see some results!

You’ve seen two different ways to run the same benchmark with Benchee now, so what’s the result and what does it look like? Well here you go:

tobi@happy ~/github/benchee $ mix run samples/run.exs
Benchmarking flat_map...
Benchmarking map.flatten...

Name                          ips            average        deviation      median
map.flatten                   1311.84        762.29μs       (±13.77%)      747.0μs
flat_map                      896.17         1115.86μs      (±9.54%)       1136.0μs

map.flatten                   1311.84
flat_map                      896.17          - 1.46x slower

So what do you know, much to my own surprise calling map first and then flattening the result is significantly faster than a one pass flat_map. Which is unlike ruby, where flat_map is over two times fast in the same scenario. So what does that tell us? Well, what we think about performance from other programming languages might not hold true. Also, that there might be a bug in flat_map – it should be faster for all that I know. Need some time to investigate 🙂

All that aside, wouldn’t a graph be nice? That’s a feature I envy benchfella for. But wait, we got this whole extensible architecture right? Generating the whole graph myself with error margins etc. might be a bit tough, though. But I got LibreOffice on my machine. A way to quickly feed my results into it would be great.

Meet BencheeCSV (the first and so far only Benchee plugin)! With it we can substitute the formatting and output steps to generate a CSV file to be consumed by a spreadsheet tool of our choice:

file = File.open!("test.csv", [:write])
list = Enum.to_list(1..10_000)
map_fun = fn(i) -> [i, i * i] end

|> Benchee.benchmark("flat_map", fn -> Enum.flat_map(list, map_fun) end)
|> Benchee.benchmark("map.flatten",
                     fn -> list |> Enum.map(map_fun) |> List.flatten end)
|> Benchee.statistics
|> Benchee.Formatters.CSV.format
|> Enum.each(fn(row) -> IO.write(file, row) end)

And a couple of clicks later there is a graph including error margins:


How do I get it?

Well, just add benchee or benchee_csv to the deps of your mix.exs!

def deps do
  [{:benchee, "~> 0.1.0", only: :dev}]

Then run mix deps.get, create a benchmarking folder and create your new my_benchmark.exs! More information can be found in the online documentation or at the github repository.

Anything else?

Well Benchee tries to help you, that’s why when you try to micro benchmark an extremely fast function you might happen upon this beauty of a warning:

Warning: The function you are trying to benchmark is super fast, making time measures unreliable!
Benchee won’t measure individual runs but rather run it a couple of times and report the average back. Measures will still be correct, but the overhead of running it n times goes into the measurement. Also statistical results aren’t as good, as they are based on averages now. If possible, increase the input size so that an individual run takes more than 10μs

The reason why I put it there is pretty well explained. The measurements would simply be unreliable as randomness and the measuring itself have too huge of an impact. Plus, measurements are in micro seconds – so it’s not that accurate either. I tried nano seconds but quickly discarded them as that seemed to add even more overhead.

Benchee tries to run your benchmark n times then and measure that, while it improves the situation somewhat it adds the overhead of my repeat_n function to the benchmark.

So if you can, please benchmark with higher values 🙂

Ideas for the future?

Benchee is just version 0.1.0, but a lot of work, features and thought has already gone into it. Here are features that I thought about but decided they are not necessary for a first release:

  • Turning off/reducing garbage collection: Especially micro benchmarking can be affected by garbage collection as single runs will be much slower than the others leading to a sky rocketing standard deviation and unreliable measures. Sadly,  to the best of my knowledge, one can’t turn off GC on the BEAM. But people have shown me options where I could just set a very high memory space to reduce the chance of GC. Need to play with it.
  • Auto scaling units: It’d be nice to for instance show the average time in milliseconds if a benchmark is slower or write something to the effect of “80.9 Million” iterations per second for the console output for a fast benchmark.
  • Better alignment for console output. Right now it’s left aligned, I think right alignment looks better and helps compare results.
  • Making sure Benchee is also usable for more macro benchmarks, e.g. functions that run in the matter of seconds or even minutes
  • Correlating to that, also provide the option to specify a warmup time. Elixir/Erlang isn’t JITed so it should have no impact there, but for macro benchmarks on phoenix or so with the database it should have an impact.
  • Give measuring memory consumption a shot
  • More statistics: Anything you are missing, wishing for?
  • Graph generation: A plugin to generate and share a graph right away would be nice
  • Configurable steps in Benchee.run: Right now if you want to use a plugin you have to use the more “verbose” API of Benchee. If Benchee gains traction and plugins really become a thing it’d be nice to configure them in the high level API like %{formatter: MyFormatModule} or %{formatter: MyFormatModule.format/1}.

So that’s it – have anything you’d like to see in Benchee? Please get in touch and let me know! In any case, give Benchee a try and happy benchmarking!