[pve-devel] [PATCH ksm-control-daemon] ksmtuned: fix large number processing

[Stefan Lendl]

Thomas Lamprecht <t.lamprecht at proxmox.com> writes:

> Hi,
>
> Am 28/02/2024 um 23:47 schrieb Roland:
>> any reason why this did not get a response ?  (i do not see rejection of
>> this  ,nor did it appear in
>> https://git.proxmox.com/?p=ksm-control-daemon.git;a=summary )
>
> No reason, but even if this looks pretty straight forward, positive
> feedback would still help to speed this up – did you test this
> successfully? Then I could apply it with a Tested-by: name <email>
> trailer.
>
>> 
>> and, while we are at ksmtuned, i think it's is broken, especially when
>> run on ZFS based installations, as it's totally mis-calculating ram
>> ressources.
>> 
>> https://forum.proxmox.com/threads/ksm-is-needlessly-burning-cpu-because-of-using-vzs-and-ignoring-arcsize.142397/
>
> Yeah KSM could definitively do with some more love, lets see if we can
> allocate some dev time for this.
>
> The RSS (which rsz is an alias for) vs. VSS (vsz aliased) looks
> interesting, and VSS really seems to be the wrong thing to look at to me
> (albeit without deeper inspection of the matter).
>
> FWIW, depending on how the sum is used it might actually make even more
> sense to use PSS, i.e., the proportional set size which better accounts
> for shared memory by dividing that part between all its users, as if
> e.g. 10 QEMU processes have 100 MB of shared code and what not in their
> RSS, using RSS one would sum up 900 MB to much compared using PSS, but
> what's the correct one here is then depending on how they result is
> used.
>
> @Stefan, as you checked this out, would you care checking out the VSS
> vs. RSS vs. PSS matter too? I.e. checking what should make more sense to
> use and actually testing that out in a somewhat defined workload.
>
> The ZFS ARC thing is something else and might be a bit more complicated,
> so I'd focus first one above at that seems to provide better
> improvements for less work, or at least with less potential to build an
> unstable control system.
>
> thanks,
>  Thomas

I agree summing up processes it would make sense to use PSS.
Unfortunately, ps does not report the PSS.

Using VSZ was introduced in cd5cf20 without further explanations.
Upstream is using RSS as well so I though it would be save to use RSS as
well, as it gives a more accurate sum than VSZ.

I will send a new version that includes reverting to RSS.
>From the awk printf perspective, I think this should work as is.
We applied the patch on an enterprise customer and no problems where
reported since.

Best regards,
Stefan