Plans for Realtime Kernel and Realtime Configuration Support in Proxmox?

[insatomcat]

Hi everyone,
I’m currently exploring Proxmox for some use cases that require low-latency and deterministic behavior - particularly in the context of real-time workloads. I’m wondering if there are any current plans or ongoing discussions around:

• Supporting a real-time kernel (e.g., with the PREEMPT-RT preemption model).
• Enabling and integrating real-time system configurations, such as:
  
  • Integrated workflows for CPU isolation, IRQ affinity, and low-latency tuning
  • vCPU pinning, and using real-time scheduling policies (e.g., SCHED_FIFO, SCHED_RR)
  • Integration or support for tools like tuned to do the kernel configuration customization options relevant to RT

Are there any official efforts or community-led initiatives in this direction? And if not currently on the roadmap, would this be something of interest to the Proxmox team or the community?

Thanks in advance, and looking forward to your thoughts!

Best regards,

 

[ReinovaTech]

Hello @insatomcat, looking for the same answers here!

With proxmox VE 9 shipping with kernel 6.14.8-2, the realtime support on the hypervisor level should be baked in and very likely should be possible to enable it as in any other Linux distro with kernel > 6.12, i guess. I am not sure what the consequences might be on the behavior of the hypervisor itself. Have you had any chance to try it out in the mean time?

 

[insatomcat]

An answer has been given on the development mailing list:
https://lists.proxmox.com/pipermail/pve-devel/2025-June/071561.html

As far as I understand, having kernel ≥ 6.12 doesn’t automatically mean realtime is enabled. What got merged was the PREEMPT_RT code, but the realtime preemption model still has to be selected at compile time. To my knowledge Proxmox ships the standard preemption model, not RT — unless they decide to provide a dedicated RT kernel, which seems unlikely based on the mailing list reply.

 

[BobhWasatch]

Even a dedicated RT kernel won't provide real-time guarantees unless the processes that want such guarantees use the FIFO or RR schedulers, which isn't the default. Which means code changes are needed to make it work. There's also this (from the thread @insatomcat linked):

for both kernel and user space the issue is that many optimizations that
improve RT performance can either change or hurt performance characteristics
for "regular" usage.

IOW, there is a trade-off of real-time guarantees vs throughput. Still no free lunches.

 

[ReinovaTech]

An answer has been given on the development mailing list:
https://lists.proxmox.com/pipermail/pve-devel/2025-June/071561.html

As far as I understand, having kernel ≥ 6.12 doesn’t automatically mean realtime is enabled. What got merged was the PREEMPT_RT code, but the realtime preemption model still has to be selected at compile time. To my knowledge Proxmox ships the standard preemption model, not RT — unless they decide to provide a dedicated RT kernel, which seems unlikely based on the mailing list reply.

Correct, sorry if I took it for granted that the PREEEMT_RT has to be enabled and that the application must use the FIFO or RR scheduler. Still this gives the sys admin/developer the opportunity to setup a system and perform some tests. There are companies out there like Beckhoff about to release software plc runtime for their twincat rtos based on Linux and running on Docker/LXC/Podman with Hard realtime capabilities, this tells that there’s room for this kind of solution and that we would probably need to expoilt them on Proxmox too.