[virt-tools-list] Default use of caching on raw volumes

[Cole Robinson]

On 08/18/2010 11:50 AM, Jamin W. Collins wrote:
> On 08/18/2010 09:23 AM, Daniel P. Berrange wrote:
>>
>> Previously when anyone has suggested setting cache=none for QEMU,
>> other people have appeared with test results showing that other
> 
> Care to provide references?  I've yet to find a single report where
> caching provides better performance with VMs stored on raw volumes
> except for the use of writeback which is not recommended for anything
> where data integrity is concerned.  My suggested change improves
> performance without risking data integrity.
> 
>> cache settings perform better for their workloads. There is no
>> perfect default setting for everyone. Changing the defaults just
>> shifts the non-optimal performance problem from one set  people
>> to a different set of people.
> 
> Again, please provide references.
> 
> Here is a quick (incomplete) list of references where changing the cache
> setting to "none" with raw volumes has increased the performance
> significantly.  As far as I can tell there have been no responses
> indicating that someone has actually performed and documented tests to
> the contrary:
> 
> https://www.redhat.com/archives/virt-tools-list/2010-August/msg00032.html
> http://virt.kernelnewbies.org/XenVsKVM
> http://www.mail-archive.com/kvm@vger.kernel.org/msg30425.html
> 
> Here are general recommendations against the use of caching with
> virtualization:
> http://permalink.gmane.org/gmane.comp.emulators.kvm.devel/48471
> http://xfs.org/index.php/XFS_FAQ#Q:_Which_settings_are_best_with_virtualization_like_VMware.2C_XEN.2C_qemu.3F
> 

This whole exchange is moot, since the QEMU maintainer has said he'd OK a
patch to make cache=none the default for block devices (granted others could
object, but I think it would be accepted).

> 
>> If you think QEMU's defaults are wrong then raise the question of
>> best default cache settings on QEMU lists again. Maybe you'll find
>> everyone agrees this time around, but there hasn't been agreement on
>> previous occassions.
> 
> QEMU's defaults aren't best in other areas either, but that hasn't
> stopped virt-manager from providing different defaults.  The default NIC
> unless one is specified is e1000 (based on the current man page
> information).  Yet, virt-manager (via virtinst) overrides this and
> explicitly sets the NIC based on guest specification.  The same is true
> for the guest HD.
> 

virt-manager/virtinst changes the defaults in this case because it has
information that QEMU does not: the OS being installed. However your are
asking us to change the cache value based on the same information that QEMU
has, overriding their default, without any sufficient reason why it's good for
virt-manager but not good for QEMU. If it's good for both, then the proper way
forward is to change it in QEMU.

- Cole