IO caching on xen?

IO caching on xen?

[marian klein]

Hi

I am trying to use the xvm/xen in OpenSolaris on the top of zvol.

The combination of xen with zfs might be the tricky and does not have to
deliver what I expect.
I am worried that the xen (or any other VM) with zvol device from ZFS
(or any other file system.) may remove the benefits of ZFS,
 because the guest OS can write something to disk but the xen itself
can do some buffering
or lazy writing and it may delay to write to ZFS backend.

Does it make sense to use ZFS with xen to protect the guest completely
against accidental power outage or against the broken disk/cables?

I understand there is no implicit persistence implemented  in the
xen/xvm, so the actual
state of the guest system is lost on the power outage. However I would
like  to have
a consistent file system in the guest OS stored physicaly on the
mirrored backend zfs. to protect explicitly written data.

Can someone point me to a relevent documentations so I can read more about it?

I probably need some guarantees claims from ZFS and some guarantees
claims from xen and then I can judge.

I expect you (as a scientist) have much better overview of the proper
documentation or at least the ideas on the topic the virtual machines
and what in theory can and what cannot  be guaranteed.

Will ever the persistence be implemented the way the EROS/Coyotos or L3/L4 has?

Marian Klein

Re: IO caching on xen?

[Samuel Thibault]

Hello,

marian klein, le Tue 15 Jul 2008 11:22:05 +0000, a écrit :
> I am worried that the xen (or any other VM) with zvol device from ZFS
> (or any other file system.) may remove the benefits of ZFS,
>  because the guest OS can write something to disk but the xen itself
> can do some buffering
> or lazy writing and it may delay to write to ZFS backend.

Xen doesn't do any kind of buffering or lazy writing.  It provides
barriers to the guest, to ensure any needed coherency.

Samuel

RE: IO caching on xen?

[Dan Magenheimer]

> marian klein, le Tue 15 Jul 2008 11:22:05 +0000, a écrit :
> > I am worried that the xen (or any other VM) with zvol 
> device from ZFS
> > (or any other file system.) may remove the benefits of ZFS,
> >  because the guest OS can write something to disk but the xen itself
> > can do some buffering
> > or lazy writing and it may delay to write to ZFS backend.
> 
> Xen doesn't do any kind of buffering or lazy writing.  It provides
> barriers to the guest, to ensure any needed coherency.
> 
> Samuel

I'm not sure if this is related or not...

I was party to some internal irc where it was believed
to be the case that when a file-backed vbd using file:'...'
is used, dom0's page cache may contain buffer pages that
may be inconsistent with the disk, whereas a file-backed
vbd using tap:aio:'...' does not suffer from the same problem.

Can someone confirm or deny (and if denied, was this fixed
recently)?

Thanks,
Dan

Re: IO caching on xen?

[Samuel Thibault]

Dan Magenheimer, le Tue 15 Jul 2008 12:32:53 -0600, a écrit :
> > marian klein, le Tue 15 Jul 2008 11:22:05 +0000, a écrit :
> > > I am worried that the xen (or any other VM) with zvol 
> > device from ZFS
> > > (or any other file system.) may remove the benefits of ZFS,
> > >  because the guest OS can write something to disk but the xen itself
> > > can do some buffering
> > > or lazy writing and it may delay to write to ZFS backend.
> > 
> > Xen doesn't do any kind of buffering or lazy writing.  It provides
> > barriers to the guest, to ensure any needed coherency.
> 
> I'm not sure if this is related or not...
> 
> I was party to some internal irc where it was believed
> to be the case that when a file-backed vbd using file:'...'
> is used, dom0's page cache may contain buffer pages that
> may be inconsistent with the disk

Ah, indeed, in that case the page cache is used, and the barriers may
not have proper effect.

> whereas a file-backed vbd using tap:aio:'...' does not suffer from the
> same problem.

blktap uses O_DIRECT to avoid buffering completely indeed.

Samuel