Re: [PATCH] KVM simplified virtio balloon driver

Dor Laor wrote:

On Thu, 2008-01-17 at 07:56 -0600, Anthony Liguori wrote:
  

quoted

Dor Laor wrote:
    

quoted

On Thu, 2008-01-17 at 11:25 +0100, Martin Schwidefsky wrote:
  
      

quoted

quoted

Another idea: Martin added an oom notifier to the cmm driver. Before the 
oom-killer kicks in cmm will try to free 256 pages. I think your virtio 
balloon driver should do the same - it seems to be the correct tradeoff. 
      
          

Nod, you definitly want to add an oom notifier. If 256 pages is the
correct number of pages to free is debatable. We have seen long delays
for a process that quickly eats up memory if there are lots of pages in
the balloon. The problem is that the memory management tries hard to
find memory until it decides to oom kill a process, only to be stopped
in the last second by the oom notifier. The 1MB is quickly eaten up
again so the whole things starts again. The profile of such a scenario
shows that almost all cpu is burned in the page reclaim code.

    
        

Seconded, in that case we can add a config space notification from the
guest to the host that will be triggered by the oom.
The host will get this notification and will decide whether to allow the
guest to deflate the balloon or to keep the current balloon size because
the whole host is over committed.
  
      

The host doesn't decide whether to allow the guest to deflate.  Virtual 
memory size and resident memory side are independent in KVM.  The host 
decides what the RSS is but the guest is free to determine it's VSS.  
The host provides hints to the guest about it's RSS size (via 
ballooning) so the guest can optimize it's VSS size.  If guest chooses 
to make it's VSS size to large, it will only hurt itself (by being 
forced to swap by the host).
    

ok, this is true when the new memory controller targeted for the next
kernel.In that case I more then agree.
Nevertheless, the guest should pass the host data about it's current
memory usage (working set size, oom triggering, etc) so the host would 
change the rss limit dynamically.
  

What is the host going to do if the guest is OOMing?  There is no way 
that the host can *verify* that the guest is OOMing so nothing prevents 
the guest from lying about the fact that it's OOMing in an attempt to 
get more memory.

More importantly though, the RSS limit has nothing to do with the guest 
OOMing.  Reducing the guest's RSS limit will result in the guest 
swapping, not OOMing.  Only the guest can prevent itself from OOMing by 
free'ing pages that were previously ballooned and that is what is being 
recommended here.

Regards,

Anthony Liguori

Thanks,
Dor