3 ms·
I'd hope the filesystem isn't provided from the login node, but otherwise, yes. That said, they may still run jobs with unprivileged -- I hope -- container ima
by gnufx 5y ago
I'd hope the filesystem isn't provided from the login node, but otherwise, yes. That said, they may still run jobs with unprivileged -- I hope -- container images. There are some that seem to feel the need to make life more complicated, but then you have N problems compared with stateless nodes off a networked root running basically native binaries.
- freemint 5y agoI never seen a HPC center where use of containers was encouraged. It is just running binaries from user directories on disk under different users. Attempts to introduce containers in a way which doesn't stand in the way are treated as revolutionary. Such as the Fuzzball scheduler (I found no good written source on a quick skim so have a talk) https://youtu.be/Pbmxq3dg35E https://youtu.be/Pbmxq3dg35E
- gnufx 5y agoI don't know about encouraged, but HPC-oriented container systems are certainly demanded, particularly Singularity (which I'd avoid). There was Shifter originally and now also at least Charliecloud and Enroot. The normal resource manager could do the namespacing itself, which is what you really want, but it's often possible just to run with an unpacked distribution and suitable environment variables in lieu of a different root.