NVENC & Framebuffer Sizing

-
Coming from K2 based servers, we did  not have the GPU monitoring capabilities or NVENC in our existing VDI servers.  That being said, the new monitoring features available on Maxwell and Pascal series GPUs have exposed some interesting sizing considerations when using NVENC.

As we've been putting the new Tesla P4 based servers into production, we've noticed much higher GPU usage at idle than expected.  Initially, to compensate, we've increased the framebuffer profile from "P4-1Q" to "P4-2Q" and see if that solved the performance issues and high framebuffer usage.  Indeed it did, so I went searching for why!

Turns out (and this make total sense) that the encode process is using additional video RAM over what it was using without NVENC.  Using RDanalyzer I looked at the encoder memory usage for Single, Dual, and Quad monitor systems using both CPU and GPU encode.  See the table below:


Takeaways:

  1. The NVENC vRAM/FB usage increase is approximately equal to the CPU Encoder RAM Usage without NVENC.  This would be good enough for sizing...
  2. There is a savings in CPU encoder RAM when using NVENC, but there is still RAM utilization of the CPU encoder (assuming for redirection, sound, printers, etc) especially with more monitors.
  3. Not listed, but in all of these scenarios, the CPU usage for encode goes from a full core a times to 0-2%.
As a shameless plug, this is one of the things I will be presenting as part of our vGPU and VDI story at GTC 2018 in San Jose.  Use code GMSGTC18 to save 20% off registration. nvda.ws/2iC4Bry 

Comments

  1. Koenraad W.November 22, 2018 at 8:16 AM

    Interesting read, this is one of the reasons why you can't enable NVENC on the lower end vGPU models, and you need at least 1 GB vGPU RAM.

    ReplyDelete

Post a Comment

Popular posts from this blog

Autodesk in Virtual Environments and What is "Allowed"

-
Yesterday at Autodesk University 2018, I attended a session (SD227281) that was supposed to be about "Best Practices for Virtualizing Your Autodesk Software."  What ended up happening was some sort of confusing and illogical announcement about new Terms and Conditions around what is allowed and what is not allowed in terms of licensing models for virtualized environments.  To categorize this session as devolving into an infuriating train wreck would be an accurate description. After cooling down and doing some research yesterday I am even more confused, and some of what we heard was seemingly wrong. First I'll summarize what was "announced" and what was discussed in that session by the two senior Autodesk people (John Looper and Chris Kayser). Announced in the presentation: Autodesk will permit virtualization for: - Subscriptions with single-user access - Enterprise Business Agreements - Server-based components (Think Autodesk Licensing Server
Read more

Dell R740 issues with XenServer 7.1/7.2

-
Image
My last post talked about our new Dell R740s.  Response to that post have pointed out that there are a few things worth following up with since people have been having some issues with this server/Intel Gold 61xx Processors. I had a few minor issues getting XenServer installed, but mostly around storage and networking. *I should note that my servers are showing crash dumps in XenCenter, but everything is working perfectly.  Waiting on my support case to have those SSRs analyzed (case originally started to ask about support for the H740p card in XS7.2).  I will post and update when I hear back tomorrow morning. Storage Controller - H740p We went with the H740p card which is not supported on XenServer 7.2 (although support tells me it may be soon) so we had to go to 7.1.1 (7.1 CU1) and install the driver at  https://support.citrix.com/article/CTX226401 .  If you had to load this to detect disks, make sure you install it AGAIN at the end during supplemental packs to actually have
Read more

NVIDIA GRID on Pascal - A Card Comparision

-
Image
I was all excited to post this, as I spent a decent amount of time on this subject leading up to ordering our new servers, and hadn't seen anyone else address it yet.  Then yesterday Jan Hendricks posted his thoughts on the subject which mirrored much of my findings ( https://jhmeier.com/2017/11/15/data-comparison-of-nvidia-grid-tesla-p4-and-p40/ ).  I guess that motivated me to finally finish my thoughts and get it posted as well! NVIDIA GRID 5.0 & Tesla Pascal Cards Photo Credit:  http://www.nvidia.com/object/accelerate-inference.html With NVIDIA GRID release 5.0 came support for Pascal GPUs.  While this brought many things (covered many other places), the most interesting part of this for me was that now ANY Tesla Pascal GPU works with NVIDIA GRID.  This opened us up to a wider array of options right away with 3 PCIe cards already generally available and an MXM option.  While NVIDIA pushed the P40 as the natural replacement for the M60, I wanted to look at all of o
Read more