Jump to content


Photo

Heterogeneous GPUs allowed in a single resource pool?

Started by Scott Nooner , 11 January 2017 - 07:12 PM
1 reply to this topic

Scott Nooner Members

Scott Nooner
  • 1 posts

Posted 11 January 2017 - 07:12 PM

Example:

2 nodes with qty. 2 NVIDIA GRID K2 cards

2 nodes with qty. 2 NVIDIA TESLA M60 cards

 

All other specs are identical (processor, RAM, storage, networking, etc.)

Looking to put all 4 nodes in a single pool.

 

Any technical block with this or can anybody think of any pitfalls?

 

Of course, Dell stopped selling GRID K2's in their R730 servers so....

 

Thanks!

 



Tobias Kreidl CTP Member

Tobias Kreidl
  • 18,287 posts

Posted 11 January 2017 - 11:55 PM

Scott,

 

Welcome first off to the forum!

 

It's not officially supported according to some NVIDIA notes I've seen, but evidently works. We have a server with both a K1 and K2, for example. What won't mix are K and M series in the same server, so for example, a K2 and an M10. That's because they use different licensing mechanisms and the drivers conflict with each other.

 

As far as in the same pool and with different nodes having different GRID hardware, there is currently no support for VM migrations across nodes (currently in XenServer, at least) that are tied to GRID cards, so these would of course fail if you went to a node that didn't have a compatible set of drivers on it. In off-line mode (the VM not running), you should be able to start up a VM on a different server if it can leverage the same type of GRID card.

 

If you need more info, I know all the right people to ask for details.

 

HTH,

-=Tobias