A Week of HPC Workshops
Dec. 19th, 2021 12:22 amWhen I look over my activities for the past several days it was pretty much entirely HPC Workshops in one form or another, starting with an introductory session at the National Compute Infrastructure's system, Gadi (most powerful public computer in Australia) conducted by Intersect. It was the sort of material that I am very familiar with of course, but every so often I like to drop into such courses to see how other people deliver their content. Certainly, I have more than a sneaking suspicion that they've picked up quite a bit of content from my workshops, but that's why I deliver them. To say the least, I'm not much of a fan of "hidden knowledge". The course does skip a lot of the Linux integration that I consider important, both assuming basic shell knowledge as a prerequisite and ignoring shell scripting, even when they are included in HPC scheduler scripts. On the other hand, their engagement through polls could be developed as a sort of formative assessment is welcome, as is their review of NCI's post-job metrics.
Following that workshop, the next three days I had workshops of my own to conduct, the final events of the year. As usual, I had the Introduction to Linux and HPC and Advanced Linux and Shell Scripting for HPC, and the third day was Mathematical Applications and Programming for HPC. The first two I pretty much run every month and the latter twice a year, as it alternates with quite a range of other workshops. Since the last session, it's had a pretty hefty revision, moving away from the "this is how you use R (or Octave, or Maxima, etc)" and more towards integration in job submission scripts and improving throughput and performance. This is going to be part of an ongoing trend in the coming year as well along with a stronger inclusion of Julia into the course, probably at the expense of Maxima.
It's been quite a year for the workshops, with more than 30 delivered, and something like 600 or more researchers in attendance. Of course, it is absolutely necessary and the demand is very strong and ongoing. It's one thing to leave researchers on a bit of a limb and say "read the manual", or even assume that they going to learn by osmosis (and such arguments are sometimes raised), but the scoreboard tells the story. Even if the researcher has "read the manual" (and we do put out a lot of documentation), they will always be unsure of something or find that their particular problem has been covered by the content. As a result, the University of Melbourne has ended up with a system that is heavily used (close to 100% node allocation on most days), has a very large number of users and projects, and has an impressive list of research outputs - not a week goes by without a paper being published that used the system. I really don't think we would see anything of the sort without the training workshops.
Following that workshop, the next three days I had workshops of my own to conduct, the final events of the year. As usual, I had the Introduction to Linux and HPC and Advanced Linux and Shell Scripting for HPC, and the third day was Mathematical Applications and Programming for HPC. The first two I pretty much run every month and the latter twice a year, as it alternates with quite a range of other workshops. Since the last session, it's had a pretty hefty revision, moving away from the "this is how you use R (or Octave, or Maxima, etc)" and more towards integration in job submission scripts and improving throughput and performance. This is going to be part of an ongoing trend in the coming year as well along with a stronger inclusion of Julia into the course, probably at the expense of Maxima.
It's been quite a year for the workshops, with more than 30 delivered, and something like 600 or more researchers in attendance. Of course, it is absolutely necessary and the demand is very strong and ongoing. It's one thing to leave researchers on a bit of a limb and say "read the manual", or even assume that they going to learn by osmosis (and such arguments are sometimes raised), but the scoreboard tells the story. Even if the researcher has "read the manual" (and we do put out a lot of documentation), they will always be unsure of something or find that their particular problem has been covered by the content. As a result, the University of Melbourne has ended up with a system that is heavily used (close to 100% node allocation on most days), has a very large number of users and projects, and has an impressive list of research outputs - not a week goes by without a paper being published that used the system. I really don't think we would see anything of the sort without the training workshops.