Create a performance-analysis lesson

[reid-a]

During the run-up to the ISC 25 proposal conversation, the topic of strategic plans came up.

Our existing workshop's workflow lesson ends with the scaling study, which feels like it's adjacent to some performance analysis.

From @mahermanns:

Should HPC Carpentry also develop Performance Analysis and Optimization Courses? Content from VI-HPS Tuning Workshops may help in developing a Carpentries-style course on tools?

[HPC.NRW]](https://hpc.dh.nrw/en/) has promised course material to help people assess performance problems based on performance monitoring data. Several sites in Germany use Cluster Cockpit but similar to "which scheduler to use" such content could be transferable. If HPC.NRW would directly work on material Carpentries-style, this could be a win-win situation.

This is adjacent to some of the discussions we have had with @bkmgit also.

[stderr-enst]

I will be involved with HPC.NRW's planning and organization of the course material and raised the same topic in the HPC Carpentry session today.

Performance optimization is likely an advanced topic and people often come with their own applications (and widely different issues), but having a common approach for the beginning of a workshop might be really useful.
I'm thinking of teaching about metrics, how to do reliable measurements, how to read metrics from systems like Cluster Cockpit (if available), and how to evaluate the resource utilization of HPC jobs (e.g. over-/under-used CPU, GPU, memory resources ...)

This is just some loose thinking on my end so far, but we'll start our activities in the coming weeks.
I definitely see the same opportunity for a win-win.

[bkmgit]

This is useful. However, the initial expected audience is users of existing codes. Probably the first starting point is the Chapel lesson as that has more depth for potential parallel programmers.

For interpreted languages like Python, profilers do exist, but some programming lessons are needed. Probably an introduction to parallel programming with MPI is needed (tracks for C, Fortran and Python, possibly R) first.

Other options such as Legion, UPC++, Dask, CAF are also useful, though maybe worth reaching out to respective communities.

[bkmgit]

Without doing a programming tutorial, profiling could be introduced in a where should I run the community code I use tutorial. This is a reasonable follow on to the scalability analysis.