The BoF at SC21
The BoF at SC21
The HPC Carpentry team hosted a “Birds of a Feather” session at the SC21 conference. This was a hybrid conference, and the BoF sessions were done via Zoom, with Q&A done via the Sli.do tool. Andrew Reid was on-site and hosted the on-site portion from the podium, while Trevor Keller, Annajiat Alim Rasel, Alan O’Cais, and Wirawan Purwanto were on-line, monitoring the Zoom and Sli.do questions and keeping the session flowing for all participants.
The primary purpose of this was to try to reconnect with the HPC user community, and gather feedback and guidance on how to prioritize our efforts moving forward. The BoF was a success on all counts: 20 people attended in-person in St. Louis, and 40 people attended via Zoom, with good, sometimes vibrant participation from both “factions” 😇 in the Q&A.
We imagine that the SC21 audience is probably somewhat more performance- and hardware-aware than the Carpentries community at large, and possibly also biased towards facility operators, rather than novice users. The feedback we gained is clearly valuable; nevertheless, we need to keep this intrinsic bias in mind while we evaluate and act on this input.
There were a number of survey questions planned in advance, which were presented via the polling functionality in the Sli.do. This information was captured and is available in the coordination repo. The principal results are that attendees would like to see, in descending order of preference: Future lessons on containers; Python and MPI; Dask and Julia; Ability to mix and match lesson content; Constructing half-day lessons that use the templating capability to adapting lessons to permanent on-premise facilities.
After the planned polling questions were done, a more general discussion ensued, largely captured in a CodiMD document.
A number of attendees were interested in how to give feedback on their experience with the lessons, and seemed uncertain what feedback we’re looking for. The answer is that we are mainly looking for actionable feedback¹ that helps improve the lessons for the community. The main mechanism by which we imagine this happening is through instructor and learner comments via poll (to be written), Slack, and GitHub issues. At the end of the lesson, following usual Carpentries practice, organizers should ask learners for positive and negative feedback, on green or red sticky notes, and conduct a post-workshop survey. This info, particularly the red stickies, is what we’re looking for, and ideally would be translated into issues on the appropriate lesson repository.
We will be digesting the feedback received over the next few weeks, and look forward to working through the generated issues over the coming months to ensure our lessons benefit from this excellent community participation.
Highlights
- One way HPC Carpentry can add value, beyond the lessons themselves, is as a clearing house for HPC educational resources generally.
- The lessons and resources we choose to focus on are a signal to the community about our priorities, so we should take some care in selecting these.
- MPI, implemented with low-level programming languages (C, C++, Fortran), is still the most commonly-used framework for parallel programming. It’s appropriate for HPC Carpentry to develop lessons that focus on this foundational material.
- New users are increasingly unfamiliar with the command line and file system hierarchy. Extra instructional effort may be needed to help bridge this gap, including the effort of developing GUI-focused lessons based on Open OnDemand or JupyterHub.
- Some users still struggle with the difference in character between distributed and shared HPC resources, and more familiar laptop or workstation resources. HPC is not just a big laptop.
- Interest in cloud-based HPC resources is surprisingly high, driven by people with possibly-transient HPC workloads but without institutional HPC facilities.
- “Actionable feedback” may seem flip, but we mean it sincerely: comments that include some tangible suggestion for improvement are tremendously helpful.