Yeah a large portion of research computing standardized on Red Hat (see e.g. Scientific Linux). The stability is quite important when trying to run ancient/niche scientific code that sometimes would only support (or even build on) Red Hat, and when you have a large number of nodes that need to be essentially bug-for-bug identical you want the package churn and update cycle to be kept to an absolute minimum.
The licensing of real RHEL never could have made sense in the HPC space, and I'd be shocked if a meaningful number of deployments would be moved to purchase RHEL now.
When I was a "sysadmin" in this space I always kind of personally preferred Debian, which has similar longevity to its release cycle, but it could never gain much traction.
I hope at this current juncture the HPC community might rethink its investment in the RHAT ecosystem and give Debian a chance.
> a large portion of research computing standardized on Red Hat (see e.g. Scientific Linux). The stability is quite important when trying to run ancient/niche scientific code that sometimes would only support (or even build on) Red Hat
> I hope at this current juncture the HPC community might rethink its investment in the RHAT ecosystem and give Debian a chance.
Nowadays Nix and Guix are available, and they're fundamentally better fits for this problem. Traditional Linux package managers aren't really suited to scientific reproducibility or archival work. It would be much better to turn more towards functional package management than just swap in another distro.
> Nowadays Nix and Guix are available, and they're fundamentally better fits for this problem.
No, it's not. First of all, reproducible installations for HPC nodes is a solved problem (we have xCAT to boot, for example). However the biggest factor is the hardware we use on these nodes.
An HPC node contains an Infiniband interface, at least, and you generally use manufacturer provided OFED for optimal performance, which requires a supported kernel.
I wasn't talking about NixOS or GuixSD, or OS installation.
I was talking about tools that let you run ancient software on modern distros. With Nix and Guix you don't have to hold your whole OS back just to run 'ancient software'.
Well, if my memory serves right, this investment started with RedHat's support for CERN's ScientificLinux and snowballed from there.
Then this snowball is solidified by the hardware we use, namely InfiniBand and GPUs, plus the filesystems we use (LustreFS or IBM's GPFS) which requires specific operating systems and kernels to work the way it should.
It's not as simple as "Mneh, I like Debian more, let's replace".
While I strictly use Debian on my personal systems, we can't do that on our clusters.
The licensing of real RHEL never could have made sense in the HPC space, and I'd be shocked if a meaningful number of deployments would be moved to purchase RHEL now.
When I was a "sysadmin" in this space I always kind of personally preferred Debian, which has similar longevity to its release cycle, but it could never gain much traction.
I hope at this current juncture the HPC community might rethink its investment in the RHAT ecosystem and give Debian a chance.