Host CPU and RAM usage slowly increases while all VMs/LXCs remain stable (only host reboot resets)

[Mister J.]

Hi everyone,

Host CPU usage gradually increases over uptime (≈18% → ≈30%), and overall host memory usage increases as well (≈14.5 GB → ≈18 GB), while all VMs and LXCs remain stable.

Only a full host reboot resets the behavior. VM/LXC restarts do not.

I’ve already investigated a number of common causes, so I’ll keep this as factual and concise as possible.

Environment​

• Proxmox VE 8.4.19 x86_64
• Kernel: 6.8.12-25-pve
• CPU: Intel i7-3770 (8 threads) @ 3.90 GHz
• RAM: 32 GB
• Storage: ZFS (ARC remains minimal and does not grow with uptime)
• Workload: multiple dedicated game servers (UT2004, UT3, UT4, COD4, BF2142, Xonotic, rFactor, etc.)
• Total VM usage: ~14.5 GB RAM, ~18% CPU

​

The actual problem​

Immediately after a host reboot:

Host CPU: ~18%
Host RAM: ~14.5 GB

After ~24 hours of uptime:

Host CPU: ~30%
Host RAM: ~18 GB

Important: the workload inside the VMs does not change. There is no observable increase in:

• guest CPU
• guest RAM
• guest load
• QEMU RSS
• interrupts
• IO
• slab usage
• SUnreclaim
• softirq load
• network load
• disk load

Only the overall host CPU and memory usage increase over uptime.

What has been observed across multiple uptime cycles​

Guest and VM-related metrics remain stable while overall host CPU and memory usage gradually increase over time, including:

• guest CPU / RAM / load remain stable
• QEMU process RSS remains stable
• no sustained change in IO or IO wait
• no visible change in interrupt rates
• no slab or SUnreclaim growth trend
• no obvious scheduler / NUMA imbalance changes
• no ballooning or memory pressure signals

Only a full host reboot resets the CPU and memory drift.

​

What I’m looking for​

I’m not looking for generic troubleshooting steps such as:

• “run top/htop again”
• “check interrupts”
• “check IO”
• “maybe it’s a runaway process”
• “maybe it’s a leak”

Those areas have already been investigated repeatedly across multiple uptime cycles.

I’m looking for insights from people who have seen similar host-side CPU and memory drift where:

• VMs and LXCs remain stable
• no metrics visibly escalate
• no runaway process is visible
• only host CPU and memory usage rise over uptime
• only a host reboot resets the behavior
• modern kernel (6.8.x)
• timer/tick-heavy workloads
• long uptime

​

Specific questions​

• Are there known regressions in 6.8.x related to:
  • scheduler
  • cpuidle / pstate
  • KVM halt polling
  • NO_HZ / tick handling
  • virtio
  • io_uring
• Would it make sense to test:
  • an older PVE kernel branch (6.5 / 6.2)
  • a newer kernel (if available)
  • forcing CPU governor to performance
  • disabling io_uring for VM disks
• Are there known cases where:
  • host CPU and overall memory usage rise over uptime
  • guest-side metrics remain stable
  • only a host reboot resolves the behavior

Any input is appreciated. I mainly want to know whether this is a known pattern with the current kernel stack, or if I’m hitting an edge case.

Thanks in advance.

PS.
The situation a few hours after a reboot of the host:





After a few days:





A reboot after a few days:

 

[Johannes S]

It's most likely caching see https://linuxatemyram.com and https://forum.proxmox.com/threads/f...ram-usage-of-the-vms-in-the-dashboard.165943/

Tldnr: It works as designed