[sheepdog-users] cluster snapshot save and load speed

[Liu Yuan]

On Tue, Jan 21, 2014 at 09:56:14AM +0100, Valerio Pachera wrote:
> 2014/1/21 Liu Yuan <namei.unix at gmail.com>
> 
> > How did you start sheep? is '--nosync' sepcified? If not, the low speed is
> > imposed by slow 'sync' request handling of disk.
> >
> 
> Yes, I forgot to mention it
> sheep -n /var/sheep /mnt/sheep/dsk02 -c zookeeper:192.168.2.44:2181,
> 192.168.2.45:2181,192.168.2.46:2181 -i host=192.168.10.4 port=3333 -w
> size=200000G dir=/mnt/sheep/dsk01

You can specify G, T, P for size, so in your case size=200T will be more human
readable.

> 
> 
> > Loading of snapshot might be slower because it is more disk intensive (it
> > need
> > generate and store parities for EC or extra copies for replciation).
> >
> 
> Does a write request of a guest differ from a snapshot load?

No, but how they issue write request will make difference for performance.
Currently, QEMU's sheepdog driver has a high performance async request hanlding,
which I think is better than that in 'cluster load', from your test.

You can verify how bad the IO performance of dog, 'vdi write' is much slower
than qemu-img convert.

This means there is much room for future enhancement.

> I know they both generate the same result: n copies or parities.
> I don't know if they do it the same way.
> BTW, as shown in another mail, snapshot load uses the default -c 3.
> I still think 6M/s is slow. I get not less than 30M/s writing from a guest
> with -c 3.

I'll fix it soon. Thanks for your report.

Thanks
Yuan