[Gluster-users] Low Performance Problems

[Stephan von Krawczynski]

On Thu, 18 Jun 2009 17:29:35 +0200
Martin Reissner <mreissner at wavecon.de> wrote:

> Hello Stephan,
> 
> did a quick test with the same fileset only replaced distribute by
> replicate on the two servers. This was with the GlusterFS patched Fuse
> though. Filesystem cache was flushed on all boxes between the tests.
> 
> write: 127s
> read: 106s
> 
> Compared with the distribute results replicate seems to perform worse.
> Here's the distribute results on the exact same setup again:
> 
> write: 90s
> read:  60s
> 
> Martin

Hi Martin,

this is very interesting. Obviously fuse needs work. I really expected
that ;-) Comparing your old values with the patched fuse version shows this
clearly. 
The above numbers are interesting, too. In the write case you can see
that 127s-90s=37s seems to be the time the network stuff needs to send the
double data size of replicate vs. distribute.
So one can deduct that 90s-37s=53s are burnt down by fuse+glusterfs client.
That makes 58% of the total time which sounds a lot.
Unbelievable is the difference in reading, though. Replicate should take
nearly double the time distribute takes for reading, that sounds voodoo.

Regards,
Stephan
 
> Stephan von Krawczynski wrote:
> > On Thu, 18 Jun 2009 15:02:34 +0200
> > Martin Reissner <mreissner at wavecon.de> wrote:
> > 
> >> [...]
> >> NFS:
> >> write: 74s
> >> read:  36s
> >>
> >> GlusterFS 1 Server:
> >> write: 332s
> >> read:   59s
> >>
> >> GlusterFS 2 Servers with Distribute:
> >> write: 331s
> >> read:   60s
> > 
> > Can you produce the same test for replicate, too?
> > A really interesting setup for people who want to get rid of NFS...
> > In theory the minimum time (caused by a FE network) should be below 23s for
> > read and write or maybe 46s for replicate write case. I know you have GBit
> > ethernet, but your disks won't cope with that anyway, so one would be content
> > with a factor 2 in real life. Nevertheless your test really shows that NFS is
> > not that bad.
> > A local-disk FUSE fs would be an interesting comparison, too.
> 
> 
> _______________________________________________
> Gluster-users mailing list
> Gluster-users at gluster.org
> http://zresearch.com/cgi-bin/mailman/listinfo/gluster-users