KVM Page Cache vs Proxmox ARC

[andy77]

Hi @ all,

I am experimenting with KVM Linux Guests (Debian) and asking myself if it really makes sense to have Linux guest OS use Page Cache at all? Because, as I understand, ARC should do the exactly same stuff directly on the node, right?

But how to disable it at all in the guest, as sysctl vm.pagecache seems to be removed form newer kernel.

Or do I maybe think wrong, and the "standard" behaviour makes sense?

Thanks for any hint
Regards

 

[LnxBil]

ARC implies ZFS and yes, the ARC is separate from the "normal" block cache for filesystems on your hypervisor due to implementation details on the solaris compability layer (spl) that comes with ZFS.

For VM caching vs. Hypervisor caching: you will always have the problem of caching multiple times and the cache "nearer" to your application is always better. Yet I do not know how to disable it in your VM completely. If you need that, you may consider moving your application to an LX(C) container, which is better at this, but still uses the memory as often as it is needed due to cgroup limitations.

We can only hope for memory dedup inside of the hypervisor kernel, which will solve this and many more problems. We already have this with kernel same page merging (KSM), but only asynchronously and time-based, but still working.

For ZFS your will also have no benefit regarding storing the same block multiple times if you are not using deduplication, which will lead to major other problems, if you do not have enough ARC.

 

[andy77]

Hi LnxBil,

thank you for the detailed answer.

 

[spirit]

For me, it's better to have page cache on guest, on don't use page cache on host (cache=none for vm disk).
If you do live migration, you don't loose the page cache.