To be honest the only thing which wonders me is, why some people have downplayed the issue when many others showed enough evidence. And that includes also Proxmox stuff. If you do not know what I mean, read the whole thread.
High VM-EXIT and Host CPU usage on idle with Windows Server 2025
- Thread starter orange
- Start date
Do you know if there is a way to confirm if AVIC is working in the guest?
This is something I was trying to figure out over recent days, I can only find ways to confirm its loaded in amd_kvm module and if its present on host CPU, but not ways to confirm its active in the guest.
It happened because of the usual fanboyism and the fact that a lot of Proxmox users are running homelabs, not production-grade enterprise setups that need to work across multiple continents. Those of us who actually make our living doing this work and have to produce results knew by page 3 of this thread that this was a serious problem.
It doesn't matter what the metrics said - the VMs felt slow to end users, and that's all that matters. Well-tuned virtualized infrastructure feels exactly like bare metal, and anyone who says otherwise doesn't know what they're doing.
For the people who were dismissing us, their node / nodes / clusters probably sit 50% idle all the time or more. However, like I said to one of these yahoos back in March, I live in the real world. My clients live in the real world. I can't tell them, "Well, Proxmox developers are waiting on the QEMU team to get their shit together and develop a fix, so we just have to sit tight and probably December 2026 to March 2027, it'll be all smoothed out."
The amount of people here who have a strong understanding of enterprise virtualization density, socket thermals, and core C-states is less than I think both you and I imagine. I can understand and appreciate the predicament the Proxmox development team is in - they have to wait for someone else to fix things so they can integrate those fixes into their product. Annoying. Been there. However, that's not an answer my clients will accept.
For most of my clients that have highly mixed VM configurations, we've switched them to Hyper-V on Windows Server 2025 Core for all Windows VMs on their newer hardware and for all their Linux / BSD VMs, we keep them running PVE 9.2 on older hardware. I sincerely hope the fix from the QEMU team comes soon though, I would much rather recommend PVE to clients, and will do so in the future once performance is at parity - or close enough to it - with Hyper-V / ESXi.
I come from software engineering and I understand that there are usual *ways* how to handle problems and if the problem is of a complex manner in a 3rd party component (like Qemu is to PVE), I accept that waiting for upstream to fix is actually the best strategy. What I do not accept is, when the Vendor does not thrive to provide all necessary information to upstream (did they even create a bug report in upstream Qemu now?) or worse just says that everything is ok (on his infrastructure, but that makes no difference). That is ignorant to all the hard work some people have put in to pinpoint and distill the problem. That's acting on the level of a mega corporation, where reputation is built through press departments, sports sponsorships and scholar system. Proxmox you are not there yet, so get back down to ground and start listening when somebody works for you for free!
Last edited:
Does the registry hack mitigate the problem also on your vSphere or you did not try?
Then it's still possible that it's a completely different issue, so you should give it a try (just for research, even when you would never use it in production).
well i had 2 test then.. and result is the same in vsphere.. spike shows reboot after adding key.. this is on fresh installed 2025 datacenter edition ( dashboard edition)
memory was less as well..
memory was less as well..
Last edited:
So the only hypervisor not showing this behaviour is now HyperV?
And can you really imagine the scale of this, how much energy is wasted world-wide?
And can you really imagine the scale of this, how much energy is wasted world-wide?
Last edited: