THreadripper 2990wx not using all cores...

Njen · March 23, 2020, 07:41:51 PM

I have two Threadripper 2 2990wx computers (217.6 Ghz each) and a dual Xeon E5-2630 v4 computer (88 Ghz). All three have 64gb of RAM.

When I render a scene I have built on all of the computers, the Threadrippers are averaging between 40% to 60% CPU utilisation, and in some moments, I have seen the utilisation go down to almost 20%. The dual Xeon is always at 100%. I'm using Deadline to render with no UI, and the Threadrippers are completing renders only around 20% faster. I would expect the Threadrippers to be over twice as fast, or is there something I missing?

WAS · March 23, 2020, 07:49:46 PM

I believe this is simply because Terragen isn't currently NUMA aware. I believe the 2990wx is 32core 64 thread. So I believe Terragen is only utilizing 32cores/threads.

Njen · March 23, 2020, 08:33:25 PM

More observation: it looks like in some parts of the image the utilisation is 100%, but in many other parts, it is lower. I have a hunch it might be the v2 cloud layer where it starts to slow down. I will do more tests.

WAS · March 23, 2020, 08:44:08 PM

That could very well be. If some buckets slow down, while other move on through and complete, utilization will drop. Finished buckets won't won't allow working buckets extra help.

Njen · March 23, 2020, 09:31:59 PM

My buckets are small enough where I am getting like 30% utilisation, and the render is only half way done, with at least a third of the image that hasn't started yet rendering. This lack of utilisation is quite worrying.

WAS · March 23, 2020, 10:13:36 PM

See here: https://planetside.co.uk/forums/index.php/topic,27266.msg271001.html#msg271001

Njen · March 24, 2020, 04:35:23 PM

Thanks for that link. Based on the information on that page, I found the coreprio utility that was designed to help the NUMA issue, and it helped. Now I'm seeing between 80% - 95% utilisation.

WAS · March 24, 2020, 04:58:24 PM

Quote from: Njen on March 24, 2020, 04:35:23 PMThanks for that link. Based on the information on that page, I found the coreprio utility that was designed to help the NUMA issue, and it helped. Now I'm seeing between 80% - 95% utilisation.

Wow, yeah that's a whooole lot better. Glad you're getting better performance. I still haven't fixed why my servers second Xeon isn't being properly recognized.