contents for section "Performance aspects of CernVM-FS"

[boegel]

No description provided.

[boegel]

To assess startup performance, we can look into a single-binary as a base case (HPL?), a typical scientific app (OpenFOAM), and a large Python app (TensorFlow).

For OS jitter, OpenFOAM is a good use case, since it's known to be quite sensitive to OS jitter (if just one core is temporarily busy with something else than running OpenFOAM, the whole multi-node run will be significantly slower).

[HereThereBeDragons]

if you dont have a benchmark setup yet, i would have some python scripts.. but they need a bit of cleanup first.
eventually we want to use them in some form to have a continous performance testing.
but let me know if you'd prefer using those or use your own solution

[ocaisa]

Right now we don't have much, performance at different scales (getting the files into the cache, and then from cache to load) is only something we've just started to look at.

As regards getting data into the cache, we have a bash script that measures the performance of our public S1's: EESSI/eessi-demo#24 . We are keen to see what kind of performance CDNs can deliver for us.

From cache to load, we have nothing, so anything you have would be a welcome starting point.

[HereThereBeDragons]

i opened up a PR with a cleaned up version of the performance benchmark i used for chep: cvmfs/cvmfs#3372
while for sure it will still change, e.g. accepting command line arguments, it is already now fully functional. if you already want to run know some benchmarking tests.
all you need to modify are the user params in start_benchmark.py and start_visualization.py (and read the comments there)
all code expects to be run from cvmfs/test/performance-benchmark
workflow:

1. python3 start_benchmark.py --write to--> ./data/<maybe_some_subdir>/*.csv
2. ./data/<maybe_some_subdir> --takes--> python3 start_visualization.py --writes-to--> ./results/<maybe some subdir>/*.pdf
   (you can set the outdir to something else if you want)

[boegel]

I have puzzled together some stats on startup performance of:

• Python (python -V): https://github.com/boegel/cvmfs-tutorial-hpc-best-practices/blob/perf/cvmfs_perf_python_hpcugent.ipynb
• TensorFlow (python -c 'import tensorflow'): https://github.com/boegel/cvmfs-tutorial-hpc-best-practices/blob/perf/cvmfs_perf_tensorflow_hpcugent.ipynb

Still work-in-progress, because not all tests were done with the same Python/TensorFlow versions, but the results look pretty good, seems like they will definitely support the narrative we have in mind for this section.