[sheepdog-users] 答复:sheepdog data write performance with 5 node

Gerald Richter - ECOS richter at ecos.de
Mon Nov 25 09:06:55 CET 2013


Thanks, one more question:

Does the qemu cache setting (i.e. cache=none,writeback,unsafe,directsync,writethrough) has any impact on sheepdog cache handling and the usage of the object cache?

Regards

Gerald

> -----Ursprüngliche Nachricht-----
> Von: Liu Yuan [mailto:namei.unix at gmail.com]
> Gesendet: Montag, 25. November 2013 08:58
> An: Gerald Richter
> Cc: sheepdog-users at lists.wpkg.org
> Betreff: Re: [sheepdog-users] 答复:sheepdog data write performance with
> 5 node
> 
> On Mon, Nov 25, 2013 at 08:49:42AM +0100, Gerald Richter - ECOS wrote:
> > >
> > > I have added an entry on this topic
> > >
> > > https://github.com/sheepdog/sheepdog/wiki/Why-The-Performance-
> Of-
> > > My-Cluster-Is-Bad
> > >
> >
> > Thanks for adding this, it is very helpful. One question is left:
> >
> > Which combination of cache/journal/nosync makes sense. For highest
> performance does it make sense to use all three, or might it better to just
> use cache+nosync, because journal causes extra disk writes (cache to
> overcome network latency and nosync to get better write performance on
> the server).
> >
> > Which combinations of these three make sense?
> 
> I think
> 
> 1 object cache + nosync (like sheep -w size=200G -n, best when host
> memory is big enough)
>   object cache + nosync (like sheep -w size=200G,directio -n)
> 
> 2 object cache + journal (like sheep -w size=200G -j size=8G)
> 
> makes sense and you'll experience better performance with 1 than with 2.
> And 'nosync' | journal duplicate each other, so setting either one is good
> enough.
> 
> Thanks
> Yuan




More information about the sheepdog-users mailing list