A couple other thoughts: It may be helpful to test each portion of the architecture independently to see what the performance characteristics are i.e. a single server's local disk, single NFS server/client, single OrangeFS server/client, etc. From there you may be able to better understand if and where the performance issue is occurring.
The behavior of the NFS exported OrangeFS filesystem may not be what you want. I don't have a list of the particular semantics that cause problems but the OrangeFS VFS mounted file system doesn't have the semantics NFS expects from a local file system and we've seen some odd behavior in the past when testing it. Maybe others have done more thorough testing with NFS on OrangeFS and there is a configuration with a known set of limitations. Thanks, Michael On Mon, Jun 20, 2011 at 12:30 PM, Emmanuel Florac <[email protected]>wrote: > Le Mon, 20 Jun 2011 18:08:01 +0200 > belcampo <[email protected]> écrivait: > > > Do I expect too much ? I thought it should scale more or less lineair. > > Yes, first NFS-export goes through the filesystem interface, which > isn't very fast. It's actually quite surprising that you've got such > high performance, because last time I checked NFS over PVFS was quite a > dog... :) -- well to be frank I didn't really test it since version > 2.8. > > Then you forgot how data flows in such a cluster. When you're pulling > 100 Mb/s via NFS, you're necessarily inducing 50 Mb/s of cluster trafic > between the nodes (actually a bit more). So you're basically > dissatisfied getting only with 150% of the maximum network > throughput? ;) What did you expect, ponies? :) > > I don't know of any cheapo 100Bt interface actually able to pump 100 > Mb/s full duplex (or really 200 Mb/s)... > > > netbook 1 netbook 2 > __ _ PVFS 50 Mb/s __ _ > [__]|=|<------------------[__]|=| > /::/|_| /::/|_| > | > | > | NFS 100 Mb/s > | > | > v > ____ __ > | | |==| > |____| | | > /::::/ |__| > > > > To get better performance in such a setup, you need at least two nics > per node : one for NFS data and one dedicated to cluster communication. > > regards, > -- > ------------------------------------------------------------------------ > Emmanuel Florac | Direction technique > | Intellique > | <[email protected]> > | +33 1 78 94 84 02 > ------------------------------------------------------------------------ > > _______________________________________________ > Pvfs2-users mailing list > [email protected] > http://www.beowulf-underground.org/mailman/listinfo/pvfs2-users >
_______________________________________________ Pvfs2-users mailing list [email protected] http://www.beowulf-underground.org/mailman/listinfo/pvfs2-users
