Frequent CPU stalls in KVM guests during high IO on host

[GETRON]

@gkovacs i will thank you so much,
because, this options and restrictions helped me to stabilise my Proxmox host server, VM to.

The host server run now stable, about 50% less RAM memory usage, and run solid, no more random restarting, and, what important, i can now run sheduled backups of VM from proxmox - no more freezing of VM in a middle of backups.

Great, because, i have nowhere found this information, for this options, like your proposed, and it WORKS GREAT.

Thank you so much, this was so important to me.

Reagrds from Switzerland.

 

[GuiltyNL]

We aggressively limit the ZFS ARC size

You did this to limit maximum memory usage, right?

Wouldn't it work to limit the 'vm.dirty_ratio' option? Because that is the maximum percentage of memory your system may use before flushing it to disk.

On our Node it has been limited to 20%. I see nice flushes every time he uses 20% of our RAM.

I think the standard setting was something like 60% and that could cause using all the free memory the system has left.

 

[gkovacs]

You did this to limit maximum memory usage, right?

Wouldn't it work to limit the 'vm.dirty_ratio' option? Because that is the maximum percentage of memory your system may use before flushing it to disk.

On our Node it has been limited to 20%. I see nice flushes every time he uses 20% of our RAM.

I think the standard setting was something like 60% and that could cause using all the free memory the system has left.

The issue described in this thread (KVM processes and kernel freezing while host is doing heavy disk or network IO) is somewhat mitigated if the vm.dirty_ratio is set to the minimum, like 2 or 1. Any value above that causes the error to appear very frequently.

Limiting the ZFS ARC size has nothing to do with this issue AFAIK, it's only required for the stability of the server, so that ZFS will not compete for memory with the KVM guests and the host kernel page cache (as it could cause unstablility).