3 ms·
I also work in HPC, and our current cluster runs Ubuntu. We're a bit of an outlier though, as you say CentOS is the most common platform. A lot of the ML/DL st
by jabl 6y ago
I also work in HPC, and our current cluster runs Ubuntu. We're a bit of an outlier though, as you say CentOS is the most common platform.
A lot of the ML/DL stuff is developed on Ubuntu first, causing occasional headaches when deploying on CentOS, similar to traditional HPC except the other way around. Oh well..
As ML gets more popular, and now this, maybe there's a chance of breaking the RHEL (rebuild) monopoly in HPC.
- bayindirh 6y agoOh, hi! It's interesting to see clusters running on different operating systems. Seeing Ubuntu on the scene is nice. From what can we see, our users can run their ML/DL loads relatively effortlessly on CentOS for some time. Moreover, with GPU support on Docker, we guess that people will migrate to containers to run their loads so distribution will become irrelevant in the future. Do you use IB on your nodes? If yes, is it easy to deploy the kernel modules?
- jabl 6y ago> From what can we see, our users can run their ML/DL loads relatively effortlessly on CentOS for some time. Yes, containers are a god-send for those apps that require some specific version of a specific distro. > Do you use IB on your nodes? If yes, is it easy to deploy the kernel modules? Yes, we do. We use Mellanox OFED, which supports Ubuntu, so it's no more difficult than on CentOS. Perhaps even slightly easier, since minor Ubuntu version updates don't break the kernel internal ABI's so there's no need to wait for all your out-of-tree kernel stuff (MOFED, Lustre, whatnot) to be updated before you can upgrade.