Grid K2 performance difference between pass-through mode and vGPU

Virtual GPU Forums General Discussion
1

I was running a benchmark test FurMark on Grid K2 in both pass-through mode and vGPU mode. The performance difference I found between them is quite high.

Eg: In pass-through mode, on 1 physical GPU I get a score of 2594.
But in vGPU mode (running alone), the score is 1714. Incidentally this score is same for all types of vGPUs (K280Q (highest allowed), K260Q, K240Q).

I have 2 questions regarding this:

  1. Is the performance difference justifiable between pass-through mode and highest allowed vGPU (K280Q)?
  2. Performance of all the vGPUs is same when run alone, does this mean that the vGPUs utilize the maximum capability of the GPU if possible and are not limited by their type? Although the frame-buffer sizes are fixed for each.

Consequently when I deploy 2 vGPUs (K260Q) on same physical GPU and run the benchmark together, the score is halved (861) which is justified.

I’m using Xenserver 7.0 as the hypervisor and XenCenter to manage and run the VMs. Guest OS used were Windows Server 2012 R2 and Windows 8.

2

Hi Ambuj,

did you disable the FRL for the vGPU profiles? Sounds to me that you compare Passthrough with K280Q and FRL enabled.

Regards

Simon

3

Hi Simon,

I did not know about the FRL. However I just tried to disable it and reboot the VM and run the benchmark but I’m getting the same score as earlier (1714).

I did this on the Hypervisor shell for the UUID of my VM:

xe vm-param-set uuid=00607ac8-6486-faab-7526-e435859c6573 platform:vgpu_extra_args="frame_rate_limiter=0"
5

@mcerveny @sschaber

What can we conclude form this then? What could be the plausible reasons for the performance difference?

Thanks and Regards,
Ambuj

6

Hi Simon,

Can you please update me on the possible reasons for the performance difference as I need to conclude this from my side.

Thanks,
Ambuj

7

Hi Ambuj,

I don’t have a good explanation other than Furmark probably doesn’t handle vGPU very well. Never ran Furmark before therefore no experience here. Can you try another benchmark? SpecviewPerf 12 for example?

Regards

Simon

8

Regarding the vGPU profile results, increasing the frame buffer will not increase the performance of the GPU, nor is it sposed to.

Example; if you were playing a game and your configured settings maxed out the frame buffer of your GPU (quality = maximum, resolution = 4k, texture detail = ultra etc etc) then increasing the frame buffer will allow better visual performance (if you were maxing it out). But the performance of the GPU will still be the same. It’s still running at the same speed, but the game will play a lot more smoothly. The same is true of professional graphics, the higher the settings, the more monitors you add into the system, the more frame buffer you will need to deliver the desired experience.

I wrote a very basic post on benchmarking and some basic system tuning options here: https://gridforums.nvidia.com/default/topic/993/grid-vgpu-benchmarks/maximising-performance-using-up-to-date-drivers-and-basic-system-tuning/

It’s mainly around Operating System power settings and driver versions, although there are hardware power settings, BIOS configurations (including cooling policies) and hypervisor power settings to consider as well.

As I said, it’s very basic, more directed at newcomers to the GPU world. However, one of the things I touch on in there, is that once you’ve bought your hardware, your chances of being able to increase its performance are limited, and unfortunately, increasing frame buffer (in the context you are testing with) is not one of them. You’re into software refinement, optimizations and overall system tuning to squeeze a bit more performance out.

If you want more performance, you either need to purchase a faster GPU, or you’re then into the world of overclocking (in that case, check out the GeForce forum) but this is an unlikely option, considering you have a K2 :-)

As Simon mentions above, check out SpecviewPerf for more relevant, complete system benchmarks.

Regarding the performance difference between Passthrough and vGPU, with the K1 / K2 GPUs, there were some features that were only available through Passthrough, and not vGPU. I don’t have a complete list of these, but I believe CUDA used to be one of them. Maybe a lack of these settings is contributing to the difference in performance?

Regards

Ben

1 Like