Anyone who’s had to deal with an unfamiliar operating system knows that you can have the fastest platform in the world, but if the operating system isn’t there for you, it might as well be a million-dollar conversation piece. IBM had a long history of in-house operating systems. Old timers might remember CP/M, and in the 1990s, there was OS/2. In the server space, AIX was king, gradually displacing every competing UNIX implementation for enterprise workloads. For years, if you got a Power system, it was probably running AIX.
In the HPC world, however, we switched to using Linux long before many other industries. It was serviceable enough, free, accepted community contributions, free, enough like UNIX to run the same applications at scale with minimal source code tweaking, and oh, did I mention it was free? While the rest of the HPC world made this transition, IBM kept plugging along with AIX. AIX and Linux are different enough that switching from one to the other is like switching from Spanish to Italian. They’re related, sure, but definitely not the same. By 2014, when I got the chance to try out a BlueGene/Q system called Mira, AIX had become so unfamiliar to HPC users that scientists were struggling to get their source code up and running on the system. Everything was different in AIX, even familiar Linux debuggers like gdb weren’t available. It’s not a knock on AIX; what had become a mainstay of enterprise infrastructure was just too different for HPC, and later, AI.
Everything changed in 2019 when IBM acquired Red Hat. They saw the way the broader IT industry was evolving with Linux, and today, Red Hat Enterprise Linux is a pillar of IBM infrastructure. They’ve even got LinuxOne mainframes! IBM isn’t locking people into the Linux they own, either, as SUSE Linux Enterprise Server is fully supported on Power. What this means is not only that RHEL and SUSE are fully supported on Power, but so are the downstream derivatives. Even though IBM formally converted the CentOS project to an upstream variant of RHEL, Gregory Kurtzer stepped in and delivered the ideal free solution for AI and HPC with Rocky Linux, which today is the most popular Linux in the AI-HPC space. Actually, it’s not just #1 in the Ai-HPC space; Rocky is the most popular enterprise Linux distribution, period.
Here at Tachos High Performance Computing, we’ve been putting a specially-configured Power 10 server with 120 cores (960 threads) and 2 TB of OMI-DDR5 4000 (64 total channels for 2 TB/s of bandwidth) using Rocky Linux. The amazing thing is just how uneventful the entire experience has been. All the packages just work. I can use my favorite open-source tools, like gcc, gdb, gprof, Open MPI, and more. Here are a few of the applications we’ve built and run in Rocky Linux for Power, never having to change a line of source, and the majority of build scripts working as-is:
- OpenFOAM
- LAMMPS
- GROMACS
- SU2
- Open RADIOSS
- WRF
- HPCG
- FinanceBench
That’s hardly an exhaustive list. As we work through various applications for clients and partners, what we’ve found is that Rocky Linux, with all its tools and features, is an excellent platform for enabling the compilation of your applications for Power, meaning spending less time fumbling with code and more time getting work done.
The ease of working with Rocky Linux on Power makes this incredibly well-engineered platform more accessible for AI-HPC than it’s ever been. We’re looking forward to trying out even more of the CIQ stack, like Fuzzball and Warewolf, to see for ourselves just how flexible and straightforward running HPC and AI applications on Power is today.