Re: [Qemu-devel] Re: [PATCH] Add cache=volatile parameter to -drive

[Anthony Liguori]

On 05/26/2010 08:06 AM, Avi Kivity wrote:
> On 05/17/2010 03:58 PM, Anthony Liguori wrote:
> > On 05/17/2010 05:14 AM, Alexander Graf wrote:
> > > Usually the guest can tell the host to flush data to disk. In some 
> > > cases we
> > > don't want to flush though, but try to keep everything in cache.
> > >
> > > So let's add a new cache value to -drive that allows us to set the 
> > > cache
> > > policy to most aggressive, disabling flushes. We call this mode 
> > > "volatile",
> > > as guest data is not guaranteed to survive host crashes anymore.
> > >
> > > This patch also adds a noop function for aio, so we can do nothing 
> > > in AIO
> > > fashion.
> > >
> > > Signed-off-by: Alexander Graf<address@hidden>
> >
> > I'd like to see some performance data with at least an ext3 host file 
> > system and an ext4 file system.
> >
> > My concern is that ext3 exaggerates the cost of fsync() which will 
> > result in diminishing value over time for this feature as people move 
> > to ext4/btrfs.
>
> In fact, btrfs is currently unusable for virt because O_SYNC writes 
> inflate a guest write to a host write. by a huge factor (50x-100x).  
> cache=writethrough is 100% unusable, cache=writeback is barely 
> tolerable.  As of 2.6.32, cache=volatile is probably required to get 
> something resembling reasonable performance on btrfs.
>
> Of course, we expect that btrfs will improve in time, but still it 
> doesn't seem to be fsync friendly.

So you're suggesting that anyone who uses virt on btrfs should be 
prepared to deal with data corruption on host failure?  That sounds to 
me like btrfs isn't ready for real workloads.

Regards,

Anthony Liguori