Does terragen supports NUMA nodes ???

Started by Fluss, March 06, 2018, 04:41:50 am

Previous topic - Next topic

Fluss

Hi,

I'm currently testing the free version to see if Terragen could be part of my workflow. The only complaint I have is that Terragen is using only one core of my workstation. I tried to override core count with no luck. On High core count systems, Windows is creating processor groups called numa nodes. Does Terragen support it ?

My system :
_win 10 pro x64
_dual xeon 2696V4 total : 44cores/88threads
_64Gb RAM

Matt

At the moment there is an issue on these high core-count systems. We hope to fix this in an update soon. However, you can suggest Terragen to use more threads than it detects automatically, and the OS will distribute those threads without Terragen needing to be aware of the hardware. One way to do this is to enable an override on the Startup panel in Preferences (set the desired number of threads, although it calls them cores). Thats a per-machine (per user) setting. If rendering on a network where you want to control the number of threads for each computer there's a command line option, e.g. "-threads 88". As a last resort if you need to force the number of threads on many computers but can't change the command line, it can be set in the render node on the Advanced tab. There you can set both min and max to the number you want. This is stored in the project (scene file).

Bear in mind that TG might not be as efficient at 88 threads. In some cases it might be slower than at 44. This can depend on the contents of the render.

Matt
Just because milk is white doesn't mean that clouds are made of milk.

Matt

Actually, regardless of what other steps you take, you'll need to increase the max from the default 64 in the render node's Advanced tab. So you can either set both min and max to 88 (if this project will only be rendered on this computer), or set only the max to 88 while leaving the min at 0 (to make the project portable to other computers) and then set the core detection override in Preferences to 88.

Matt
Just because milk is white doesn't mean that clouds are made of milk.

Fluss

If I do that, I got a prompt that too many threads are set coz it's running only on one CPU (so half the threads). We were able to change some BIOS settings on Xeon V3 to allow software that does not support numa nodes to take advantage of all cores, sacrificing some performance. But with Xeon V4, this setting has gone so we're condemned to use only one core unless the software support NUMA nodes... That's quite annoying that most software does not take care of that, especially in CGi industry which relies heavily on computation performances... Looks like only renderers are supporting it ATM which is a shame IMHO.

Fluss

Quote from: Matt on March 07, 2018, 06:02:04 amAt the moment there is an issue on these high core-count systems. We hope to fix this in an update soon.
Any news on this issue?

Matt

This is on the roadmap for v4.5. I'll start work on it soon.
Just because milk is white doesn't mean that clouds are made of milk.

lemon_mirage

Greetings,

I'm new to Terragen. I have a 64core/256Gb node, and Terragen is using only a maximum of 50% of my cores/threads. Is it an actual limitation on Terragen ?
I tried the startup overide and to put 128 threads in the renderer. Nothing helps.

Best,

Olivier

WAS

Quote from: lemon_mirage on October 01, 2021, 07:01:37 amGreetings,

I'm new to Terragen. I have a 64core/256Gb node, and Terragen is using only a maximum of 50% of my cores/threads. Is it an actual limitation on Terragen ?
I tried the startup overide and to put 128 threads in the renderer. Nothing helps.

Best,

Olivier
Yes, Terragen is not NUMA aware. Split your scene up between two instances of Terragen (crop render).

crisb

Hello lemon_mirage,

This limitation is something that we're working on. It's not as simple as it sounds though.

cheers,

cris.

WAS

Curiously, maybe you can answer @crisb . Is it possible to have a divided sequence? Like, it only renders every other frame by a defined offset or what not, so you could split the task? That may help as well and be less to set up?