I just added a P4 to one of my hosts for exactly this. My servers can only handle low-profile cards, so the P4 fits. Not the most powerful of GPU's but I can get my feet wet.

From my perspective, there's literally money on the ground for any virtualization platform to pick up VDI with Intel. The GPU's are affordable and performant for VDI work. They currently work with Openshift and Proxmox is at work on it.

When IBM/RedHat "killed" CentOS, the rest of the world took a hint and left. Companies and projects left CentOS in droves as the future of their products were in jeopardy due to the loss of CentOS.

At this point, the damage is done.

These latest 8.3 update speeds are still slower than a 13 year-old Xeon E3 1230.

@Andrew Thanks for that added detail.

Your success to Wasabi is encouraging. Perhaps Planedrops performance issues with BackBlaze B2 is related to a specific combination of implementation of S3 between BackBlaze and XO.

Things to test:

XO to AWS
XO to Wasabi
XO to BackBlaze

Theoretically, the performance should be the same to all S3 endpoints.

@olivierlambert

From what it looks like...If XCP-ng could move to kernel 6.8+ the Intel SR-IOV option should be on the table. That's a pretty massive jump from 4.19 and I have no idea how much effort that would be for you guys.

Nothing fruitful to add....

But...

Oooof....

This will be somewhat messy to clean up. I'm rooting for you guys though!!

I'm attempting to virtualize some media functions and doing video work on raw CPU is just crazy. (I've got an app running 80% on second gen 3Ghz Epyc with 12 vCPU's). I'm pretty sure if I had a vGPU to throw at it I could cut that down significantly.

From what I've seen, it looks like AMD is broken, NVidia works but requires expensive hardware and equally expensive licensing, and Intel isn't even remotely going to work.

Is that pretty much the state of things?

Thanks!

James

@olivierlambert While VDI is maybe not as vital as it once was...I'm experimenting with multimedia work in XCP-ng. Having a VM with GPU off-loading of CODEC encoding would be nice. It's a pretty big CPU hit to make that go.

@olivierlambert Ideally you need to be somewhere into Kernel 6. 6.12 is sticking out in my head, but I'm not positive when support got fully integrated.

@olivierlambert If I remember right, you should be able to see 62 VF's on that card. There might be a tool needed to define how many VF's are present like on a NIC.

@Forza said in Epyc VM to VM networking slow:

Would sr-iov with xoa help backup speeds?

If you specify the SR-IOV NIC, it will be wire-speed.

@olivierlambert From Intel's page:

"With up to 62 virtual functions based on hardware-enabled single-root input/output virtualization (SR-IOV) and no licensing fees, the Intel Data Center GPU Flex 140 delivers impeccable quality, flexibility, and productivity at scale."

https://www.intel.com/content/www/us/en/products/details/discrete-gpus/data-center-gpu/flex-series.html

Generally speaking...Just because you haven't had direct requests doesn't mean the feature isn't desired.

It's easy enough to look over the current features and because it's not listed, you assume it's not there and move on to find the next viable solution.

Proxmox and OpenShift seem to be killing it in this space with Intel Flex GPU's.

I can't believe that there's not much interest.

Anyway...As far as I know these are pretty well supported in newer kernels. I think you need to be fairly deep in kernel 6. Given that XCPng/XenServer is currently running on kernel 4 with a bunch of backports, this might be a little problematic.

Last I knew, Intel is not charging licensing fees for using vGPUs like NVidia does.

With a working driver in XCP-ng and no licensing fees, this is could be a real cost-effective VDI platform.