rdma-core VS ohpc

Great news! Soft-RoCE interface has been successfully created by using a very simple command instead of recompile whole kernel.

As a part of my semester thesis I need to get a ~5 year old RDMA benchmark to cooperate with the about one year old RDMA core libraries that got put into upstream Linux in the mean time.

Name : rdma-core Version : 34.0 Release : 1.fc33 Architecture : i686 Size : 115 k Source : rdma-core-34.0-1.fc33.src.rpm Repository : @System From repo : updates Summary : RDMA core userspace libraries and daemons URL : https://github.com/linux-rdma/rdma-core License : GPLv2 or BSD Description : RDMA core userspace infrastructure and documentation, including initialization : scripts, kernel driver-specific modprobe override configs, IPoIB network : scripts, dracut rules, and the rdma-ndd utility. Name : rdma-core Version : 34.0 Release : 1.fc33 Architecture : x86_64 Size : 120 k Source : rdma-core-34.0-1.fc33.src.rpm Repository : @System From repo : updates

We have plenty of licensed RHEL, but in isolated environments the hurdle of connecting to a Satellite server or their subscription hub on the internet is too high -- at least with Rocky and the ilk available. For this set up, the licensing model doesn't match reality, at least not easily.

Are we really going to build out compatible configuration management, monitoring, logging, etc? -- it's not a seamless transition. How much time do we have to put towards this?

And yes -- there is software compatibility issues. Look at the OpenHPC software distribution, it's designed for SUSE or Enterprise Linux: https://github.com/openhpc/ohpc/wiki/2.X

I recommend you just stick with HPC centric tools are workflows. Your scientists aren’t going to learn k8s as you said. SLURM is the scheduler you want and if you’re new to HPC, I recommend taking a look at https://openhpc.community

the general consensus is that pam_slurm_adopt is the better module (that's just one dude's opinion but his citations are good) - the advantage is that not only will it gatekeep SSH access, it'll also drop their SSH session into the cgroups that are constraining the user's resource limits, which also means their CPU usage will show up in sacct for the job (if the user has multiple jobs running on a node their ssh session may get dropped into the wrong one, no help for that)

Check this: https://openhpc.community/ (this helped me a lot when I started. I'm no longer the admin of such systems)

Overall, if you're already in a RedHat-based environment, an installation of OpenHPC is pretty straightforward. Their reference implementation assumes you have a head node for the scheduler that all other nodes NAT through, but that's not a 100% requirement as much as a common setup. It also assumes you can reformat the compute nodes and dedicate them to HPC work, so if you need to keep the systems available as normal workstations, you'll need to deviate a bit. You could also use the OpenHPC instructions as a guide for what packages to install, but it may take longer to get everything right.

https://github.com/openhpc/ohpc/wiki/1.3.X Newer versions of OpenHPC don't seem to releasing XCat guides anymore unfortunately.