[Gluster-users] Performance problems

[Olivier Le Cam]

Aurélien ROUGEMONT wrote:
> will this be ok for a web servers farm ?
> what dies it mean ? What is the most common bs= equivalent for a www 
> server ? Do i have to change the block size of the underlying filesystem ?
> 
> (by the way, i've managed to get the same results with an opensolaris 
> 2009.06 + glusterfs 3.0.3)
> 
> I would be glad to get some more explanations about this
> 
> Thank you all for taking the time to answer to my wonderings.

I face myself to the same interrogations. In addition to the results 
with "dd" posted before, I have done some benchmarks with the "postmark" 
tool this morning, which is relatively easy to simulate webserver 
accesses to the filesystem with. I'm also a bit dubious regarding the 
GlusterFS performances in a web farm environment.
Here are some comparaisons from the same client to both an NFS (v3) 
server and a "RAID-1" GlusterFS cluster (postmark was set with "1000 
base files" and for "50000 transactions"):

As GlusterFS Client, with all default translators enabled:
	- 200 transactions per seconds
	- data read: 642 kilobytes/s
	- data written: 670 kilobytes/s

As NFS Client
	- 961 transactions per seconds
	- data read: 3.10 megabytes/s
	- data written: 3.24 megabytes/s

(see details at http://docs.google.com/View?id=dg4cv5kw_74ptxfqqgq)

Important notice: the NFS server used here is in production and under a 
quite heavy load. We should expect even better results (by far) if that 
NFS server were 100% dedicated to this benchmark!

I'm now wondering if GlusterFS is indeed a good choice for a webserver 
farm filer despite of its very interesting features in term of 
reliability and scalability.

Like Aurélien, I'd appreciate your comments on this topic!

Thanks in anticipation and best regards,
-- 
Olivier