stop ksm on the fly?

[Tom Horsley]

I see ksmd using 60% cpu on my kvm host which is running
many different totally different linux distros as KVMs.
It seems likely to me that ksm will be of no value on this
machine, and I could use the cpu time for the VMs themselves :-).

My question is this: Can I stop ksmd without breaking the
existing VMs? What is the best way to turn it off
"on the fly"? (Or is "service ksmd stop" perfectly
safe)?

[Dan Kenigsberg]

On Fri, Dec 18, 2009 at 08:18:35PM -0500, Tom Horsley wrote:
>  I see ksmd using 60% cpu on my kvm host which is running
>  many different totally different linux distros as KVMs.
>  It seems likely to me that ksm will be of no value on this
>  machine, and I could use the cpu time for the VMs themselves :-).
>  
>  My question is this: Can I stop ksmd without breaking the
>  existing VMs? What is the best way to turn it off
>  "on the fly"? (Or is "service ksmd stop" perfectly
>  safe)? 
Perfectly safe - as long as you have enough memory. Otherwise, the VMs
might start swapping to disk which may be even worse than what you have
now.

Now since there's still to be desired in ksmtune, would you give some
info about your system (total ram, number and sizes of VMs, and `more
/sys/kernel/mm/ksm/*`)

Regards,
Dan.

[Tom Horsley]

On Sat, 19 Dec 2009 15:47:36 +0200
Dan Kenigsberg wrote:

>  Perfectly safe - as long as you have enough memory. Otherwise, the
> VMs
>  might start swapping to disk which may be even worse than what you have
>  now. 
Yea, once the install I was doing finished, I went ahead and tried
it and nothing drastic happened (and I got some cycles back).

>  Now since there's still to be desired in ksmtune, would you give
> some
>  info about your system (total ram, number and sizes of VMs, and `more
>  /sys/kernel/mm/ksm/*`) 
The smolt profile is:

 http://www.smolts.org/client/show/pub_fde3037c-3bfd-4a28-9f73-601d559b6567

4 dual core opterons (total of 8 processors), 8 gig of memory, a gazillion
different VMs defines for lots of different linux distros and versions
(we use it for testing our software under a all those versions), so it
is highly unlikely that any two machines would have many similar pages.
Each VM is configured with 1 gig of virtual memory, and my testbeds never
have more than 7 running at the same time (which leaves plenty of
memory for the host as well as each vm).

Here's the ksm files (but this is long after I already disabled ksm, so
I don't know if these are meaningful anymore):

[root@xen ~]# more /sys/kernel/mm/ksm/*
::::::::::::::
/sys/kernel/mm/ksm/full_scans
::::::::::::::
16085
::::::::::::::
/sys/kernel/mm/ksm/max_kernel_pages
::::::::::::::
1024201
::::::::::::::
/sys/kernel/mm/ksm/pages_shared
::::::::::::::
44390
::::::::::::::
/sys/kernel/mm/ksm/pages_sharing
::::::::::::::
90333
::::::::::::::
/sys/kernel/mm/ksm/pages_to_scan
::::::::::::::
1250
::::::::::::::
/sys/kernel/mm/ksm/pages_unshared
::::::::::::::
815693
::::::::::::::
/sys/kernel/mm/ksm/pages_volatile
::::::::::::::
72635
::::::::::::::
/sys/kernel/mm/ksm/run
::::::::::::::
0
::::::::::::::
/sys/kernel/mm/ksm/sleep_millisecs
::::::::::::::
20