[Sheepdog] Databases, write barriers etc

[Brian Candler]

On Tue, Aug 09, 2011 at 06:59:54PM +0900, MORITA Kazutaka wrote:
> > If an app wants to update a single block, will this require Sheepdog to
> > fetch an entire 4MB object, modify part of it, and write it back?
> 
> The app sends only the data to be written, and the server daemon
> overwrites the existing object.  The app doesn't fetch the entire
> object.

Thank you. So it's just like a normal block write, and the issues are (a)
network latency, and (b) software latency and scheduling delays in the
gateway and object manager processes.

network latency would exist in an iSCSI SAN environment anyway.

I know very little about how Linux schedules KVM guests. Would it help to
run the sheep process with a negative 'nice' value?

Regards,

Brian.