[Gluster-devel] various questions AFR + unify + small files
Amar S. Tumballi
amar at zresearch.com
Tue Jun 24 21:07:12 UTC 2008
> Aim is 20 nodes (x4 cores) with 1TB disk each, on gigabit switch
> - we are interested in AFR + unify for having 10TB x2 replications
> - maybe NUFA
> - we have tons of small files (10kB) which are our read bottleneck
> - we want performance (currently we easyly exhaust 4 NFS servers)
> - our users won't do concurrent write to the same file, except append
> to logfile (we don't care of order as far as everything is logged)
> 1/ which version should we use? 1.3.9.tgz or some more recent from tla ?
1.3.9 has some minor problems (like excessive logging in posix_setdents
etc), we are about to make 1.3.10 after a week or two with afr write order
fix. But I suggest latest tla of 2.5 branch as the best bet. (as there is
only bug fixes happening on that branch).
> 2/ i guess we should use all the performance modules
'io-cache' is useful. as you told all are small files, 'read-ahead' is not
> 3/ docs seems to say we should do (unify over NUFA over AFR) ?
> or is it better to keep it simple and do only (unify over AFR) ?
NUFA is used when each server has a client part too.
> 4/ i'm a bit confused about namespaces, especially if we use
> unify + NUFA + AFR. I probably missed a link explaining ns
> and how avoid single point of failure.
> Btw, http://www.gluster.org/docs/index.php/GlusterFS_Volume_Specification
> should be better advertised in documentation, as it is so important.
> I'll write my step-by-step-glusterfs-howto-for-newbies and send it if
> we succeed.
Cool! that would be a great help.
> Thanks in advance for answers, and for your great job.
> Best Regards.
> Alain Baeckeroot
> Gluster-devel mailing list
> Gluster-devel at nongnu.org
[bulde on #gluster/irc.gnu.org]
http://www.zresearch.com - Commoditizing Super Storage!
More information about the Gluster-devel