At Fri, 31 Aug 2012 08:58:22 +0800,
Liu Yuan wrote:
> 
> On 08/30/2012 10:42 PM, MORITA Kazutaka wrote:
> > I wonder if there is a situation where some users want to enable a
> > disk cache of Sheepdog, but the others don't want to.
> 
> The biggest problem is we don't have a sane disk cache implementation for 
> backend (seems impossible to me).
> Current proposed approach doesn't scale. Suppose we have 1000 nodes, and sync 
> request will be issued in
> 30 second. Because we have to broadcast sync request to all the nodes, then 
> we'll end up with
> 
>  1000^2 / 30 = 33333
> 
> for every host every second. This will definitely degrade the performance. 
> That is, O(n^2) will only apply
> to very small sized cluster. If we have an O(n) disk cache for backend, then 
> I think we don't need to disable
> it.

Sorry, what I meant is "is there any situation where some qemus want
to enable a disk cache of Sheepdog, but the others want to disable
it?".  Because we are arguing that we need to disable the cache from
qemu command line (e.g. with 'cache=none').

I know Hitoshi's patch doesn't scale, but we can disable it with a
sheep command line when using many nodes for Sheepdog.
-- 
sheepdog mailing list
[email protected]
http://lists.wpkg.org/mailman/listinfo/sheepdog

Reply via email to