Optimizing Ryzen 9 genomics workstation

Hi Everyone,

I recently put together a Ryzen 9 5900x genomics workstation, and I’m trying to take my time to set it up and get some really good performance out of it.

Currently I’m running Kubuntu 20.4 to avoid possible memory leak issues with GNOME.

I’m coming up with a list of things to optimize, and I want to make sure I don’t miss any low hanging fruit.

Currently I’m looking into:

  • undervolting the voltage/frequency curve with PBO 2
  • taking advantage of the intel Math Kernel Library and this performance workaround here (assuming things haven’t changed) to get the best speed for matrix operations. (running lots of numpy.)
  • Tuning the memory clocks/Infinity fabric ratio, and memory timings.

While tech youtubers talk about overclocks all the time, I worry that I’ll miss some important steps (like the intel MKL) that I could really benefit from.

I appreciate any suggestions!

2 Likes

I always get worse clocks and temps if I use any of the automatic overclock algorithms in the BIOS like PBO.

I get better constant overclocks and temps if I set a fixed clock ratio and a fixed cpu voltage sufficient to maintain those clocks without errors.

I run a constant load 24/7 at 94% of the threads available for scientific computing.

Hmmm, fishy.

If your workload is long running (and doesn’t checkpoint for whatever reason - please look into Apache arrow if you’re not using it already) use cgroups.

I use them somewhat hackishly to limit chrome from taking up more than 32G of ram.

Lookup cgroup-tools, specifically cgrules.conf.