Community
Arnold for 3ds Max
Rendering with Arnold in 3ds Max using the MaxtoA plug-in.
cancel
Showing results for 
Show  only  | Search instead for 
Did you mean: 

Arnold Not Using All Cores On Dual 64 Core Server (NUMA Nodes)?

9 REPLIES 9
Reply
Message 1 of 10
drinaldi
1060 Views, 9 Replies

Arnold Not Using All Cores On Dual 64 Core Server (NUMA Nodes)?

I am rendering Maya/Arnold on an AMD Dual 64 Core EPYC 7742 server. It has 256 threads, but due to the 64-bit architecture of Windows Server 2019, the cores are divided into 2 NUMA nodes. Maya and Arnold are not able to see or access these other cores - when I render, I see only 128 buckets (as well as confirmed in the log), as well as only half the 128GB of RAM. Is it possible to render using all of the available cores? It seems a shame to only utilize half the power of the server...

Tags (1)
Labels (1)
9 REPLIES 9
Message 2 of 10
thiago.ize
in reply to: drinaldi

I'm not aware of anything in Arnold that would be limiting this, especially the memory. Do you have any other applications that are able to use all the resources of your machine? In case this is a Maya limitation, can you try running a scene through kick to see if that also is limited? Can you post the top of the arnold log that has the version info and CPU and memory details?

Message 3 of 10
drinaldi
in reply to: drinaldi

Or- if possible- can you render from the command line using a flag to specify the NUMA node to utilize? Does such a flag exist?

Message 4 of 10
drinaldi
in reply to: drinaldi

The memory seems to be a give away. When I render the same scene locally on a dual 18-core (72 threads) with 256Gb RAM and no NUMA nodes ( single system), Arnold sees all the RAM and uses 72 buckets.

Message 5 of 10
thiago.ize
in reply to: drinaldi

My machines with two numa nodes don't have this problem. However, I've only ever tested Win7 and 10. Maybe there's a setting in Windows Server 2019 you need to adjust?

Message 6 of 10
thiago.ize
in reply to: drinaldi

Arnold doesn't have a flag for that. I know that in Linux there are commandline tools, such as numactl, for specifying what numa node to run on. I have no knowledge about whether Windows offers that.

Message 7 of 10
drinaldi
in reply to: thiago.ize

Hi Thiago- that’s certainly good news to hear. Can you share the specs of the machine you have that have two NUMA nodes?

Message 8 of 10
thiago.ize
in reply to: drinaldi

One is a dual socket Xeon (56 logical cores total) and another is an AMD 3990x which can be run using 1 through 4 NUMA nodes (bios option). We also have many other customers with dual socket machines. However, you're the first I've heard of using a machine with 256 threads. Do try and see if other applications are registering all the resources, because if they aren't, this suggests a possible OS setting/issue.

Message 9 of 10
o.blyznjuk
in reply to: drinaldi

we have several render nodes with the same dual 7742 and win10 pro, all cores are typically being utilized, but preprocessing (subd/disp) time in some cases is even worse than on old nodes with low-core xeons. Hovewer, that problem goes away under centos 8.

Message 10 of 10
drinaldi
in reply to: drinaldi

Hello everyone - thanks for the ideas. After quite a bit of research, it seems a limitation of the combination of the Windows code and Autodesk code. Windows will limit any processors cores over 64 (128 threads) into their own Processor Groups. These Processor Groups also have their own memory group (local vs remote memory). So for this server that is running Windows Server 2019, the OS will indeed recognize dual 64 cores and 256Gb of RAM, HOWEVER any program instance can only spawn to one of the Processor Groups. So when you launch Maya, it will ONLY see and render using one Group (ie: 64 cores/128 threads instead of the 256 the system has). Same goes for memory as well. So instead of seeing 256 buckets, I can only use 128 in any Arnold render - half the power of the system. Solution - you can either settle for half the power of your specs and wait for Autodesk/Arnold to code up access to multiple NUMA nodes in the same instance, or leave Windows, which I did. I installed Linux CentOS 7 (like Oleg said above - thanks!) and it has zero problems seeing and utilizing all cores and memory per render. Thanks!

Can't find what you're looking for? Ask the community or share your knowledge.

Post to forums  

Autodesk Design & Make Report