In the second half of 2011, I ran a Multicore Computing course (COMP8320) for advanced undergrads and postgrads. For their 3rd assignment, they wrote some simple linear algebra algorithms and then did a fairly comprehensive performance analysis based on a largish number of experiments on the SCC. The issue I had as instructor of course was a find a way that a dozen or so students could safely and fairly use the SCC at any time they wanted (especially in the last 2 days before their deadline!!) without someone having to be there to monitor and if needed reboot the machine.
Anyway, it can be done! A rather inelegant combination of scripts and lockfiles, and once I added a script that identified and purged any process on a core which looked like it was from a leftover job, it all went very well (no reboots for the last 3 days after that). Of course, I did not tell them any other way of accessing the SCC other than the modified rccerun script that I provided. No doubt there are more elegant solutions but if anyone is interested in how to make the SCC a pseudo- multiuser machine or even in the reasonably interesting performance experiments that the students did on the SCC, I have written it up and made the relevant source files available, please see: the course web pages .
If anyone else has had experiences is making the SCC available to coursework students, I'd be interested to hear.
We have a couple of students working on the SCC. They have set up a shared calendar (at google) to reserve time slots and the "w" command on the MCPC shell also helps to check if somebody else might use the machine at the moment. However, your approach is interesting, because it allows much more fine-grained access.
Are you interested in adding other programming models for many-core programming to your course? We use TACO, which provides very fast remote method invocations on global object pointers (a little bit similar to X10). It runs quite nicely on the SCC as well as normal PCs and clusters.
Thanks randolf, I would be interested in the fast RPC for the next (2013) version of the course, especially as a comparator paradigm to the message passing model. I wonder if X10 will be available by then.