Required shared filesystem?

[vsoch]

Hi! For the examples you provide, is a shared filesystem required? We are running a performance study, and looking for contenders (strong or weak scaling examples work that have some kind of FOM, even if just running time). We are planning to run on 4 to 64 nodes and looking for apps / proxy apps / benchmarks / synthetic benchmarks that could be candidates. The only requirement is that I can build it into a container and run it across nodes, and it's in Kubernetes so no shared filesystem for now. It's not impossible, just harder to do (and more costly in terms of time). Thanks!

[tskisner]

It is possible to run a pure simulation workflow without loading any time ordered data from disk. However, usually one wants to write out some results at the end. If one carefully chooses to use healpix pixelization, using HDF5 format, and disables parallel HDF5 writes, then map-domain products are gathered and written by a single process. In this way you could avoid the need for a shared filesystem and the rank zero process of the world communicator would write to some volume mounted location in order to have the results at the end.

It would be useful to know more about your particular goals / use case. For example, one could also run a synthetic benchmark that disables writing the output data products.

[vsoch]

Writing a result file at the end is OK! Many apps do that, and either I copy the file locally or cat to the terminal (so it appears in the log). The issue would be nodes (pods) requiring access to the same file. It sounds like that might be possible from your description?

We are primarily trying to demonstrate a means to deploy these kinds of apps in Kubernetes in an easy way (using helm). The set I've done already are here: https://github.com/converged-computing/flux-apps-helm/

Let me know if you have a suggested example (and params). I think a good mix of different use cases / types would strengthen the work, so we can be flexible to that.