Re: [RFC 01/10] KVM: reintroduce hc_gpa

[Avi Kivity <avi@redhat.com>]

On 06/06/2012 03:32 PM, Sasha Levin wrote:
> On Wed, 2012-06-06 at 15:15 +0300, Avi Kivity wrote:
>> On 06/06/2012 02:45 PM, Sasha Levin wrote:
>> > On Wed, 2012-06-06 at 14:38 +0300, Avi Kivity wrote:
>> >> On 06/06/2012 02:32 PM, Sasha Levin wrote:
>> >> > 
>> >> >> and benchmark results.
>> >> > 
>> >> > Is there a specific test you'd like to see? My tests were based around
>> >> > the "streaming test" you proposed last time this discussion came up, and
>> >> > runs of 'fio' in guests.
>> >> 
>> >> Those are fine, thanks.
>> >> 
>> >> btw, IIUC, a guest can have two tmem providers: the hypervisor and
>> >> zcache.  How does it choose?
>> > 
>> > The guest will try to use the hypervisor, unless:
>> > 
>> >  - 'nokvmcleancache' is passed as param - in that case cleancache will
>> > use the local zcache provider.
>> > 
>> >  - 'nokvmfrontswap' is passed as param - in that case frontswap will use
>> > the local zcache provider.
>> 
>> The guest may have more memory than free host memory, and its memory is
>> certainly faster; but I'm not sure how to integrate this better.
> 
> Won't the free guest memory be exhausted first by regular local caching
> before cleancache springs into action?
> 

It would.  But we don't know whether it would be better to compress and
store on the guest (if we have 2:1 compression ratio, we need to process
two pages to free one) or compress and store on the host (where the
backing store is more or less free, but may be smaller, and is more
expensive to access).

-- 
error compiling committee.c: too many arguments to function