Shared initialization step?

[simonbyrne]

In our old slurmci we have a separate init job that installs the packages and does a round of precompilation. This isn't such an issue with buildkite as agents have a cache that they can use (reducing the need to reinstall packages), but if we wanted to do something like this we would need a way to share the initialized cache between the downstream agents.

[simonbyrne]

One idea from JuliaCon BoF on Julia in Production: build a system image of all the dependencies, and invalidate only if the Manifest changes

[simonbyrne]

We could even do this into a Singularity container?

[jakebolewski]

I think that would work as long as nothing had to be written to the singularity container during runtime (after build step)

[simonbyrne]

I think singularity containers are immutable

[simonbyrne]

How about this:

• We have two shared depots, one for each configuration (e.g. /groups/esm/buildkite/depot/cpu and /groups/esm/buildkite/depot/gpu)
• We have 1 special agent (e.g. central-init) of which there can only be one instance running at a time (either via --dependency=singleton, or even simply run the agent on the login node), that is only used for the buildkite-agent pipeline upload steps.
• This agent is the only one able to write to the shared depots: during init, we have a command that instantiates the repository with both depots, e.g. something like

  module load openmpi/4.0.4
  JULIA_DEPOT_PATH=/groups/esm/buildkite/depot/cpu julia --project -e 'using Pkg; Pkg.instantiate(); Pkg.precompile()'
  module purge
  module load openmpi/4.0.4_cuda-10.2 cuda/10.2
  JULIA_DEPOT_PATH=/groups/esm/buildkite/depot/gpu julia --project -e 'using Pkg; Pkg.instantiate(); Pkg.precompile()'
  module purge

  • I checked, and it does look like we can instantiate and precompile a CUDA-aware MPI on a non-GPU node.
• all subsequent jobs add this depot to the stack, with their top depot on scratch (which should typically remain empty), i.e.

  export JULIA_DEPOT_PATH="$(dirname "$BUILDKITE_BUILD_CHECKOUT_PATH")/.julia:/groups/esm/buildkite/depot/cpu"

[jakebolewski]

If we could do this with a singleton slurm job instead of as extra logic in the agent I think that would be preferable as it is a bit more flexible. Is this possible with slurm?

[simonbyrne]

on second thoughts, I agree an explicit extra step would be better.

[simonbyrne]

Is this possible with slurm?

Yes, you give it a unique job name (say --job-name=buildkite-init), and then use --dependency=singleton, and it will only run one instance at a time.

[jakebolewski]

Ok I think this should be straightforward then

[simonbyrne]

This might be a nice solution: https://github.com/JuliaCI/DepotCompactor.jl/