keskiviikko 12. heinäkuuta 2017

Is every build system using Ninja just as fast as every other?

One of the most common arguments against Meson is that "it is only fast because it uses Ninja rather than Make, using any other Ninja build generator would be just as fast". This is always stated as fact without any supporting evidence or measurements. But is this really the case? Let's find out.

For testing one needs a project that has both CMake and Meson build definitions. I'm not aware of any so I created one myself. I took the source code of the Mediascanner 2 project, which is using CMake and converted it to use Meson. This project was chosen solely based on the fact that I wrote the original CMake definitions ages ago so I should have a fairly good understanding of the code base. The project itself is a fairly typical small-to-medium project written in C++ with a handful of system dependencies.

Compiling and running the project on a regular laptop gives a fairly straightforward answer. Both build systems are roughly as fast. So, case closed then?

Well no, not really. The project is small and machines today are very fast so a similar result is not very surprising. For a better test we would need to either convert a much larger project or use a slower machine. The former is not really feasible, but the latter can be achieved simply by running the tests on a Raspberry Pi 2. This is a fairly good real world test, as compiling programs of this size on a raspi is a common task.

The measurements

The tests were run on a Rasberry Pi 2+ running Jessie with a sid chroot. The CMake version was the latest in Sid whereas Meson trunk was used. Each measurement was done several times and the fastest time was always chosen. If you try to replicate these results yourself note that there is a lot of variance between consecutive runs so you have to run them many times. The source code can be found in this repository.

The first measurement is how long running the first configuration step takes.

CMake takes 12 seconds whereas Meson gets by with only four. This is fairly surprising as CMake is a C++ executable whereas Meson is implemented in Python so one would expect the former to be faster. Configuration step is run seldom, so it's ultimately not that interesting. Most of the time is probably spent on a full build, so let's look at that next.
CMake takes 2 minutes and 21 seconds to do a full build. Meson is 31 seconds, or 20%, faster clocking at 1 minute 50 seconds. Both systems build the same files and they have the same number of build steps, 63, as reported by Ninja. Finally let's look at the most common task during development: incremental compilation. This is achieved by touching one source file and running Ninja.

In this case Meson is almost 50% faster than CMake. This is due to the smart link skipping functionality that we stole from LibreOffice (who stole it from Chromium :). This improvement can have a big impact during day to day development, as it allows faster iteration cycles.

Conclusions

Ninja is awesome but not made of magic. The quality of the generated Ninja file has a direct impact on build times. Based on experiments run here it seems that Meson performs consistently faster than CMake.

3 kommenttia:

  1. How is Meson's generation of Xcode and VStudio files? Last time I checked it was very sketchy.

    VastaaPoista
    Vastaukset
    1. Visual Studio is working and usable but not super awesome. Xcode is wonkier.

      Poista
  2. Visual Studio should be 1-1 ninja, minus a few known bugs. The GNOME module particularly does not work with the Visual Studio backend.

    We run (almost) all the tests that are run on the Ninja backend with the VS backend too. The only ones we don't are those that are platform-specific.

    XCode backend is quite broken because there's no automated testing for it, but should be easy to get into shape for basic usage and enable it on the CI. Someone needs to spend a day or two on it. I'd be happy to guide whoever wants to do that.

    VastaaPoista