We are seeing that PDG isn't leveraging all cores of the threadripper 3990x (64cores/128HT).
When I run a basic liquid sim in just SOPs everything is great all cores are engaged.

Am I right in thinking that because PDG is build on python it's limited to a single NUMA node somehow?

I can't get more than half of the nodes engaged in processing. In this particular case it's COP's processing.
No simulation, so it could be COP's, but I doubt it.

Can we assume if we start a second houdini session it will automagically pick the least active NUMA node, or will it always default to the first?

Thanks for any info!
-Johan

What version of Houdini are you using? 18.0.527 has a fix that may be related.

I'm running this on 18.0.532.

So it's PDG and COPs with ROP Composite outputs mainly and a small commandline utility at the end of processing to set some metadata in the JPG output.
We tried to turn of SMT and some other settings, but no performance gains, so we left it as default with 2 numa cores.

For PDG, if you have 128 work items running at once, are you saying that they appear to only be using one numa cpu group?

I put the scheduler on different settings, also 64 and more then 64, but it only ever burdens the first numa node. I also ran a second houdini instance with a simple flip sim (but very small particle separation). And that just ran over all 128 cores no problem. If you want me to test something just let me know. But so far I was unsuccessful to have PDG use everything.
Here's a screenshot of the numa node CPU burden, I have enough work items I think. But if you have a test file or something I'd be happy to run it.

Could you try going to Task Manager, Detail Tab and RMB on Houdini and select ‘Set affinity’, are there any deselected cores?

I found it too, 19.5.368 still really slow on 3990x