July 5th, 2007 by lucas
With the development of rebuildd and the fact that several people are interested in re-using my scripts, I feel the need to explain how this stuff works.
First, Grid’5000 is a research platform used to study computer grids. It’s not really a grid (it doesn’t use all the classic grid middleware such as Globus). Grid’5000 is composed of 9 sites, each hosting from 1 to 3 clusters. Inside clusters, nodes are usually connected using gigabit ethernet, and sometimes another high speed network (Myrinet, Infiniband, etc). Clusters are connected using a dedicated 1OG ethernet network. Grid’5000 is in a big private network (you access it through special gateways), and one can access each node from any other node directly (no need for complex tunnelling).
Using Grid’5000 nodes
When you want to use some nodes on Grid’5000, you have to use a resource manager to say “I’d like to use 50 nodes for 10 hours”. Then your job starts. At that point, you can use a tool called KaDeploy to install your own system on all the nodes (think of it as “FAI for large clusters”). When KaDeploy finishes, the nodes are rebooted in your environment, and you can connect as root. At that point, you can basically break the nodes the way you want, since they will be restored at the end of your job.
Running Debian QA tasks on Grid’5000
None of that was Debian-specific. I will now try to explain how QA tasks are run on Grid’5000. The scripts mentioned below are in the debcluster directory of the collab-qa SVN repository.
When the nodes are ready, the first node is chosen to play a special role (it’s called the master node from now on). A script is run on the master node to prepare it. This consists in mounting a shared NFS directory, and run another script located on this shared NFS directory to install a few packages, configure some stuff, and start a script (masternode.rb) that will schedule the tasks on all the other nodes.
masternode.rb is also responsible for preparing the other nodes. Which consists in mounting the same shared NFS directory, and executing a script (preparenode.rb) that installs a few packages and configures some stuff. After the nodes have been prepared, they are ready to execute tasks.
To execute a task, masternode.rb connects to the node using ssh and executes a script in the shared directory. Those scripts are basically wrappers around lower-level tools. Examples are buildpackage.rb, and piuparts.rb.
Now, some specific details:
- masternode.rb schedules tasks, not builds. Tasks are commands. So it is possible, in a single Grid’5000 job, to mix a piuparts test on Ubuntu and an archive rebuild on Debian. When another QA task is created, I just have to write another wrapper.
- Tasks are scheduled using “longest job first”. This doesn’t matter with piuparts tests (which are usually quite short) but is important for archive rebuilds: some packages take a very long time to build. If I want to rebuild all packages in about 10 hours, openoffice.org has to be the first build to start, since building openoffice.org takes about 10 hours itself… So one node will only build openoffice.org, and the other nodes will build the other packages.
- I use sbuild to build packages, not pbuilder. pbuilder’s algorithm to resolve build-dependencies is a bit broken (#141888, #215065). sbuild’s is broken as well (#395271, #422879, #272955, #403246), but at least it’s broken in the same way as the buildds’, so something that doesn’t build on sbuild won’t build on the buildds, and you can file bugs.
- I use schroot with “file” chroots. The tarballs are stored on the NFS directory. Which looks ineffective, but actually works very well and is very flexible. A tarball of a build environment is not that big, and this allows for a lot of flexibility and garantees that my build environment is always clean. If I want to build with a different dpkg-dev, I just have to:
- cp sid32.tgz sid32-new.tgz
- add the chroot to schroot.conf
- tell buildpackage.rb to use sid32-new instead of sid32
- Logs and (if needed) resulting packages are written to the NFS directory.
Comments and questions welcomed :)