One node, multiple GPU distribution issue


#1

Hi there,

so we are rendering with redshift standalone and we have one node with multiple GPU. We want to distribute incoming jobs on each GPU on that node.
I tried the following options:

  • Slave options of that node: Concurrent Task Limit override: 2 (2 GPUs on that machine)
  • GPU affinity: Activate GPU 1 and 2.
  • redshift-options: Activate only one GPU.

However, during rendering only GPU 0 is used due to the fact that in the rendering-call of the redshiftCmdLine it says only to use -gpu 0

How can I make it happen that I use both GPUs in parallel for two jobs at the same time?


#2

use multiple slave instance, then assign each slave an affinity with a gpu.

SlaveInstance1 = GPU_Affinity=gpu0
SlaveInstance2 = GPU_Affinity=gpu1

and so on, I use this for up to 8 cards, usually in 4x2 affinity

https://docs.thinkboxsoftware.com/products/deadline/10.0/1_User%20Manual/manual/multiple-slaves.html
(I think you need to enable launching new slave instances in user groups under menu items)


#3

I’ll try that and report. Thanks!!


#4

I guess that worked, thank you!
So, but how many licenses would I need for a machine running 4 GPUs? Only one?


#5

Yeah, just one per physical machine.


#6

Great, thank you all for your input! Very good.