Re: Poor KVM guest performance on an HP rack server

[Avi Kivity <avi@redhat.com>]

On 03/12/2010 07:18 PM, Ozan Çağlayan wrote:
> Jeremy Fitzhardinge wrote:
>    
>> On 12/13/09 02:12, Ozan Çağlayan wrote:
>>      
>>> Hmm, that's because we're using/developing a Linux distribution which
>>> doesn't have 64-bit support yet. But since the guest will be a
>>> compile-farm for our needs, it's worth trying to run an 64-bit distro on
>>> the host to see the difference, thanks for the tip but how much should
>>> this affect the performance?
>>>        
>> 32-bit PAE is marginal above 4G of memory and the limit is at about 8G.
>> Above that, so much lowmem (<~1GB) is occupied by page and other kernel
>> structures that there's barely any memory to do any real work.  If this
>> is what's happening, you'd expect to see poor host performance, and slow
>> guest performance would be secondary to that.
>>      
> Updates on the issue after 3 months:
> ------------------------------------
>
> The host is now running an x86_64 2.6.32.9 kernel with kvm-2.6.32.7 built as an external module.
> The guest is 32-bit 2.6.32.9.
>    

You're running an old kvm on a new kernel.  It's simpler and better to 
run the kvm that came with 2.6.32.9.

> I've booted the guest with -smp 4 -m 4096
>
> and I'm seeing the same slowness that I've mentioned in this thread during CUPS configuration.
> I'll try to send kvm_stat output in this week and try the same configure on another guest OS to see if it differs or not.
>
> I'm also thinking to build an optimized 2.6.33 for that rack server to see if things get better or not.
>    

Better to find out what's wrong first and build kernels later.

-- 
Do not meddle in the internals of kernels, for they are subtle and quick to panic.