Fork me on GitHub

How we keep the lights on

Posted on February 9, 2020 by Troels Henriksen

I have written much about how Futhark is designed and how the compiler is implemented. In this I will switch things up a bit and instead write about how we have built our project infrastructure - from automatic compiler testing to the web server hosting this blog post.

In the same way that Futhark’s small niche affects the language and compiler design, our infrastructure is similarly impacted. We assume that whoever maintains it is incompetent, easily distracted, and not very diligent. I know that this is a fair characterisation, because I am the one who maintains it. Everything we built must therefore be highly robust, require almost no ongoing maintenance, and be very cheap (the money itself is not the biggest problem, but payment setups are always a potential failure source). Overall, while our setup is not perfect, it seems to work pretty well, and some of it might serve as inspiration to other small language projects.

Our guidelines

We have two main guidelines for our infrastructure:

  1. Avoid stateful servers and daemons that may crash or corrupt their state, and prefer batch mode programs that frequently restart from scratch.
  2. Prefer as few systems as possible.

Following guideline 1, we try to make heavy use of “serverless” cloud services, in particular Travis CI (I know I’m abusing the “serverless” term here). The batch jobs produce static files that are eventually just served by a simple web server.

Following guideline 2, all of those static files are served by the futhark-lang.org server, which is the exact same system that runs my personal website, my IRC bouncer, my outgoing email proxy, and a bunch of other things. The hypothesis here is that I am very likely to quickly find out if this system fails. I do not trust myself to set up reliable monitoring, but I do trust myself to notice when I cannot get on IRC or send email. The system itself is a fairly cheap VPS hosted by TransIP, and it runs OpenBSD for reasons I have explained elsewhere.

I am well aware that putting everything on a single system is not best practice, but best practices are meant for people who are skilled at their job, and as mentioned earlier, I am not a skilled systems administrator. While I could no doubt eventually figure out how to set up some containerised infrastructure on AWS, I would probably have forgotten how it works by the time something breaks and I have to fix it.

Continuous integration

Futhark is hosted on GitHub, and we use Travis CI as our primary CI mechanism. On every commit and pull request, a Travis job checks whether the compiler still compiles, and runs our test suite. The test suite is executed through a program called futhark test. While originally intended only to test the compiler, we now also use it to test Futhark programs and libraries on their own.

Travis jobs, unfortunately, are limited to 60 minutes at most, and building the compiler from scratch - including its own dependencies -and running the test suite can take longer than that, especially on the small machines that Travis makes available for free. Our solution is to break the build into multiple stages, each of which has its own 60 minute cap. We use one stage to build dependencies, one to build the compiler itself, and one to run the test suite, and use Travis’s support for build caches to preserve build artefacts across stages. All this was rather painful to set up, as I find Travis’s documentation rich on examples and poor on fundamentals, but it works fairly well for now (our .travis.yml).

When we release a new version of Futhark, it is also a Travis job that builds the release binary tarball and uploads it to GitHub and our web server. This is done simply by pushing a Git tag of the format vX.Y.Z.

To hedge our bets, we also run CI on Azure Pipelines. It tests only a little more than the Travis job, but without all of the “stage” complexity, as the Azure timeouts are long enough that everything can be built from scratch on every commit. While I find the Azure Pipelines web UI to be significantly more clunky than Travis’s, the setup was otherwise quite straightforward.

Both Travis and Azure run on Linux. Since we also want to support Windows, but none of the Futhark developers actually run it, we also have a somewhat bare-bones AppVeyor setup. We have had quite a bit of trouble with this one being flaky, so perhaps we will migrate Windows builds to Azure eventually.

Benchmarking

All of the preceding 3(!) CI systems hare the same flaw: they run on arbitrary cloud hardware with no GPUs. This means they cannot test the most interesting aspect of the compiler, namely how fast the generated code runs! Therefore we also maintain a Buildbot setup that runs on known hardware physically located in the university data centre, with servers running RHEL and primarily maintained (at least at the OS level) by the IT department.

Buildbot is a CI tool that is written and configured in Python. It is less popular than Jenkins, a similar tool, but I greatly prefer configuration through a python API to clicking around in a user interface. Our configuration looks like this.

The architecture of Buildbot is straightforward. A central master server listens to new commits in the Git repository, and schedules various jobs in response. These jobs are handled by workers, which in our case are GPU-equipped university machines that connect to the master. One nice aspect of this design is that only the master needs to be publicly accessible on the Internet, while the workers can be behind NATs or firewalls without any issue. In our case, we run the master on a server called napoleon, that we inherited from the HIPERFIT project, but that is mostly for historical reasons. Since all the master does is coordinate the workers and perform a few bookkeeping jobs, we should probably migrate it to the same system that runs the web server.

So, what do the jobs actually do? They run our test- and benchmark suite using all of our parallel backends (opencl, cuda, pyopencl, csopencl), using the futhark bench tool, and upload the performance results to a central directory, with one file per commit/backend/machine combination. This means that we have historical performance data on (almost) every commit.

What we don’t currently do, unfortunately, is to automatically compare new results to old ones, to identify performance regressions. This is still done manually. We do, of course, have some tooling for performing the analysis. The simplest is compare-compiler-versions.sh, a script that takes two compiler commit IDs and for all benchmarks shows the difference in performance. This is useful enough for gauging the consequences of a single commit.

For more in-depth historical analysis and visualisation, we need a more complex tool. This is the Futhark Benchmark Dashboard, which can graph performance changes for a benchmark over time (example). Such dashboards are common for compiler projects (here’s Chrome’s). However, ours is carefully designed to require no custom server-side code at all. The dashboard is a pure client-side JavaScript application served as simple static files, reading processed data files also served over HTTP.

The raw data gathered by benchmarking is organised by commit, while the dashboard needs the data to be organised by benchmark and dataset name. This pre-processing is done by a Buildbot job that runs after every benchmark job and, as always, just dumped in a web server directory. The dashboard code itself does not need to be updated. The advantage of this is also that we have no dependency on a database, along with the maintenance complexity that would bring. The downside is that the information is not truly structured, and there is no way to query it using e.g. SQL.

I should note that as I am not brave enough for JavaScript myself, the dashboard was developed as a bachelor’s project by a group of then-undergraduates at DIKU (Bo Thomsen, Frederik Lassen, and Emil Weihe). Accustomed as they were to mainstream web programming, it took them a little while to understand why I wanted such an odd architecture, but they did eventually deliver something that has been running well, and completely maintenance-free, for years.

Package repository

Futhark comes with a bare-bones package manager that I have written about before. As with everything else in Futhark, it is built around free cloud services and static file hosting. Specifically, a package is identified by a repository on GitHub or GitLab, with a tag for each released version. Eventually, I wish to extend this so that a package can be identified by any URL pointing at an appropriately formatted HTML file containing metadata, but it has not proven necessary just yet.

Apart from a tool for downloading and upgrading packages, there must also be a central repository of packages to provide discover-ability. For Futhark, this is handled by futhark-docbot, which contains a file listing the location of every known Futhark package. Every day, an automatically scheduled Travis job then looks at all available versions of each package, and generates a set of static HTML files that list all packages, along with documentation generated by futhark-doc, and uploads the result to a central location.

Automation via regularly scheduled Travis jobs is a pretty lightweight way for free software projects to perform such routine tasks. It does mean that up to a day may pass before new package versions show up in the registry (although they can be installed immediately), but this is a small trade-off in return for not having to write and maintain any server side code myself.