Monitoring vRAM usage from ESXi.

NVIDIA Virtual GPU Forums Virtual GPU NVIDIA Virtual GPU Technology
#1

Hey guys, I’m currently POCing GRID for use with Xenapp 7.6 on ESXi 6.0/ I am using LoginVSI to simulate user load, and would like a way to see GPU usage and vRAM usage without spamming nvidia-smi -l in Putty.

I have been doing a poor man view of vRAM usage by running GPUPERF on a session while having the others log in, but I can’t log or track that really.

LoginVSI can accept any form of logging in the form of a .CSV and we use vCOPS as well (which would be my preferred method of gathering performance information) - I’d like to have time stamped output I can correlate with my user session count/CPU/Memory usage of the VM, so i can decide what k2 profile I need to run my Xenapp Servers on.

Is there anything built into the NVIDIA driver that would help me here?

Thanks!

#2

Did you try more options for nvidia-smi (like "nvidia-smi -q", "nvidia-smi -q -x"+xsltproc, "nvidia-smi dmon", "nvidia-smi -l --query-compute-apps=timestamp,used_memory --format=csv,noheader,nounits"…) ("-h" :-) ?

M.C>

#3

If you’re in XenApp and using a passthrough GPU to the XenApp VM then you have all of the metrics available in perfmon.

Or you can query the WMI provider directly.

It may also be possible to use Process Explorer to export per process usage information to .csv / .xls though I’ve never tried it.

If you’re using a GRID vGPU for the XenApp server then you will be unable to measure the GPU load for the VM, but you can still measure Frame Buffer / Graphics memory utilisation.

#4

Thanks all for the Reply.

We are using vGPU and attempting to determine how many Xenapp users can we have on a Server see the benefits of the GPU acceleration while providing the minimum amount of vRAM per VM. We are simply talking knowledge workers, so I am hoping we can fit 10-20 users on a single VM running the K220q profile or something similar. scale out the K220q profile, thats (2xK2s) 32 VM’s with 10-20 users on each…you get the idea. That density is important for my POC.

You can indeed see the GPU memory usage via nvidia-smi q or using perfmon in windows. I am using GPUperf for it during testing, which doesn’t give me any pretty graphs, but does give me an idea of the density as loginVSI tests are performed.

looks like I might be able to export Perfmon to CSV as well.

We use Edgesight, Director, VMWARE vCenter and Virtual Operations Monitor, and Solarwinds SNMP polling to track performance of our current VM’s - I guess I was hoping to find something that fit in those existing Frameworks.