Benchpark mirror

[pearce8]

Use case: Running Benchpark on a system that does not have free access to internet.

Leverage Spack mirror and Ramble mirror as appropriate.

Notes from Doug from December 2024:

• mirroring in Ramble will auto-mirror with spack
• ramble workspace mirror -h
• you can add "input" mirrors using ramble mirror -h
• You might also check out ramble deployment for isolated machines
• https://ramble.readthedocs.io/en/latest/workspace.html#workspace-deployments

[scheibelp]

Assuming you've set up a Ramble workspace:

bin/benchpark system init --dest=rzvernal llnl-elcapitan rocm=5.5.1 compiler=cce +gtl lapack=intel-oneapi-mkl blas=intel-oneapi-mkl
bin/benchpark experiment init --dest ./gromacs-rocm gromacs +openmp+rocm~cuda gpu-aware-mpi=force
bin/benchpark setup gromacs-rocm rzvernal workspace/
. ./workspace/setup.sh
ramble --disable-progress-bar --workspace-dir ./workspace/gromacs-rocm/LlnlElcapitan-040fd89/workspace workspace setup

you can then create a mirror:

ramble --disable-progress-bar --workspace-dir ./workspace/gromacs-rocm/LlnlElcapitan-040fd89/workspace workspace mirror -d file://X

then

• X/software is the mirror you add to spack: spack mirror add sources file://X/software
• X/inputs is the ramble mirror you add for benchmark inputs: ramble mirror add file://X/inputs

given the above, you can use the following workflow to effectively mirror benchpark:

1. Checkout benchpark
2. Run benchpark setup + ramble workspace setup on all experiments you want to mirror
3. Run ramble workspace mirror create

• (this has to be run for each experiment/system pair)

4. A benchpark mirror can then be created from the following:
   a. the ramble workspace mirror (from [3])
   b. the clones of Ramble and Spack from the benchpark workspace
   c. the benchpark repository
5. On the other side:

• Unpack everything, run the Ramble/Spack mirror add commands (described above)
• To run benchpark, one would have to copy the Ramble/Spack repos into ~/.benchpark/

None of these steps require modifications to benchpark to work (i.e. someone could do this right now), but some could be streamlined, and some might run into potential issues

• [2] doesn't just concretize the experiments, it also builds them
• [3] we could loop over all the Ramble experiments (e.g. if there were a benchpark mirror create command)
• [4b] could run into problems if the externals aren't the same; depending on the extent of mismatch, this could get progressively more challenging
  • Maybe all the same externals are available, but in different locations: we could add a script to rewrite the external prefixes automatically (given some mapping)
  • If not all the same externals are available, that could lead to different concretizations: users would have to manually update the source mirror (this wouldn't have to occur each time the experiments were moved, but is still a potential pain point)
• Also regarding [4b], you probably don't want to copy the full benchpark workspace:
  • the concretized environments would refer to external prefixes that may not match (however, if they do match, you may want to copy over the full benchpark workspace)

See: https://ramble.readthedocs.io/en/latest/mirror-config.html#ramble-mirrors