[Gluster-users] Fw: performance evaluation of distributed storage systems

Jeff Darcy jdarcy at redhat.com
Tue Jan 22 15:59:10 UTC 2013


On 01/22/2013 04:55 AM, Sogand Shirinbab wrote:
> I'm a phd student and as a part of my research I've compared performance
> of different distributed storage systems (Gluster, Openstack,
> Compuverde). I would like you as an expert in your product to give me
> feedback on my work. What do you think about the way we've setup the
> system? does it affects the Gluster performance?

Interesting work.  Some questions:

(1) Why was Ceph excluded?  It's readily available and has an almost 
identical object-storage interface, but no reason is given for leaving 
it out.

(2) The Atom/4GB/GigE storage nodes seem like an odd choice.  Have you 
tried testing on other kinds of platforms?

(3) More information on software versions and configurations would be 
very helpful.  For GlusterFS, there are significant differences both 
between versions and between different ways of organizing a volume 
across 384 disks.

(4) What program(s) other than SPECsfs did you use to generate load? 
The results for most of the tests seem very inconsistent with those for 
SPECsfs, especially for GlusterFS.

(5) How POSIX-compliant is the structured form of Compuverde?  Does it 
have full and proper support for things like fsync/O_SYNC, extended 
attributes, or atomic cross-directory rename?  Does it use FUSE, or 
interface to the system in some other way?

(6) Several of the statements made in section 5.3 seem inaccurate. 
GlusterFS only uses rsync for remote replication, which it seems clear 
from the rest of the paper would be irrelevant to these tests.  Also, 
it's not generally true that self-heal would actually be done from the 
proxy servers (though it could be initiated from there).  Lastly, this 
is one of the areas where version/configuration differences would make a 
huge difference in the results.

(7) The nature of the relationship between BTH and Compuverde needs to 
be more explicit.  Did it include configuration/tuning help?  Bug fixes? 
  Equipment loans or other material support?  If not, then it's just a 
matter of applying usual disclosure standards.  If so, then perhaps 
representatives from other projects (OpenStack as well) should have the 
same opportunity to make sure the results are representative of current 
best practices.



More information about the Gluster-users mailing list