Xen devel - Jun 2006 - balloon.py unaware of pages on page_scrub_list

Suppose I create a domain with a "large" amount of memory (for some
value of "large") and I set "on_crash" to be
"reboot".  Then I crash the domain.  All the domain''s memory
gets put on the page_scrub_list.  Then xend attempts to restart the domain.  The
balloon code in xend sees that very little memory is free, so it attempts to
balloon dom0 down even more.  At this point, the balloon code and the page
scrubber are racing.  The domain may fail to restart, or domain 0 may be
squeezed down farther than necessary.  If dom0-min-mem is not tuned for the
machine, this can get ugly.

I think this applies to rebooting too, not just crashing.

Would it make sense to give this information to balloon.py, and apply some
heuristics?  Perhaps xend should wait for the scrub list to go to zero before
trying to balloon.  Or maybe it should consider that memory as free, and only
try to balloon (needed - to_scrub).

The most obvious way (to me) to get that information to xend is via
DOM0_PHYSINFO.  That way, the number of free pages and the number of pages to
scrub can be gotten atomically.

Thoughts?

I volunteer to code and test this if there''s some agreement on how it
should be done.

Thanks,
Chuck



_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xensource.com
http://lists.xensource.com/xen-devel

On 19 Jun 2006, at 23:15, Charles Coffing wrote:
> Would it make sense to give this information to balloon.py, and apply 
> some heuristics?  Perhaps xend should wait for the scrub list to go to 
> zero before trying to balloon.  Or maybe it should consider that 
> memory as free, and only try to balloon (needed - to_scrub).
>
> The most obvious way (to me) to get that information to xend is via 
> DOM0_PHYSINFO.  That way, the number of free pages and the number of 
> pages to scrub can be gotten atomically.
>
> Thoughts?
>
> I volunteer to code and test this if there''s some agreement on how
it
> should be done.
Your approach sounds fine to me.

  -- Keir


_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xensource.com
http://lists.xensource.com/xen-devel

Keir,

Attached is a patch that makes the ballooning logic in xend aware of
pages that are on the page scrub list.  This fixes the case where
crashing/restarting a domain can cause dom0 to balloon more than
necessary.

I changed the physinfo dictionary in Python to be in KiB, rather than
MiB, to avoid accumulating ugly rounding errors.  I tried to avoid
changing units anywhere else.

The patch is against xen-unstable.  It applies and works against
xen-3.0-testing also (in fact that''s where I did most of my testing)
but
I don''t know if you want to apply it there.  I changed the
dom0_physinfo
in a non-backwards-compatible way.  Looked nicer to me to group the
*_pages fields together, but that''s probably not something
you''d want to
do in 3.0.2...  your call.

Thanks,
Charles

Signed-off-by: Charles Coffing <ccoffing@novell.com>


On Tue, Jun 20, 2006 at  1:20 AM, Keir Fraser
<Keir.Fraser@cl.cam.ac.uk> wrote: 
> On 19 Jun 2006, at 23:15, Charles Coffing wrote:
> 
>> Would it make sense to give this information to balloon.py, and
apply 
>> some heuristics?  Perhaps xend should wait for the scrub list to go
to 
>> zero before trying to balloon.  Or maybe it should consider that 
>> memory as free, and only try to balloon (needed -  to_scrub).
>>
>> The most obvious way (to me) to get that information to xend is via
>> DOM0_PHYSINFO.  That way, the number of free pages and the number of
>> pages to scrub can be gotten atomically.
>>
>> Thoughts?
>>
>> I volunteer to code and test this if there''s some agreement on
how
it 
>> should be done.
> 
> Your approach sounds fine to me.
> 
>   --  Keir
> 
> 
> _______________________________________________
> Xen- devel mailing list
> Xen- devel@lists.xensource.com
> http://lists.xensource.com/xen- devel



_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xensource.com
http://lists.xensource.com/xen-devel