Notes from a recent workshop

[ocaisa]

I wanted to document some feedback I have based on teaching this course online over the last 24 hours.

• I used HPC jargon busting as an ice-breaker but it fell a little flat (maybe we should have a few options for this?)
• I taught the recent merge of the ssh key generation. I found it interesting and hopefully useful for them to teach this. I used the analogy of (private) key and (public) reproducible lock...with a passphrase being like adding fingerprint recognition to the key
• Some of the jobscript snippets are not consistent (small changes in what are supposed to be the same file)
• A graphic on the different type of nodes would be helpful to help show how they are somewhat the same but serving different purposes
• Some commands appear somewhat out-of-the-blue like sinfo -n node1.int.eessi-gpu.learnhpc.eu -o "%n %c %m" | column -t with no explanation
• Short form of Slurm options are used a lot, but the long form is usually much more descriptive. For our use case these might be a better option? (I guess this is a config issue)
• The diner option works but feels a little contrived if you try to use it too much to shoehorn multiple examples into it. Having a few bullets that covers the important points from this analogy would help keep things on track
• It would nice to have an optional part about how to request GPUs
• I would agree that the a recent comment from @tkphd that the Roleplay is probably too much effort (and certainly impossible in the online world)
• It is never clearly shown that (for Slurm at least) any option in a jobscript could also/instead just be given on the command line
• A good example of why you want to know about job environment variables is SLURM_NTASKS (which allows you to use mpirun -n $SLURM_NTASKS ... in an MPI example, saving you having to remember to edit a job script in two places)
• Remote graphics recommends using ssh -Y...but it is not stated that this is insecure!
• Python 3 is no longer a useful example for when you might need module load ... (without some further discussion)
• Including the / with rsync copies the contents of the directory but scp does not have this behaviour
• I did not even try to get into the pi calculation, I used the example in Simplifying the parallel application example #362 and never even showed them the code inside. I think assuming Python knowledge is unnecessary for this lesson.
• free and other commands are introduced without explanation