I haven't tried gfs yet, but did you simply test simultaneous access to the same local disk (without gfs) by 2 separate processes: generally peformance drops because of disk seek time (the reason is mechanical) Disks always have much lower performance when shared between 2 processes on the same machine or 2 processes on separate machines (it is the same as having a dual cpu)

This is just an idea... not very sure if this is your problem.

Jure Pečar wrote:

Hi all,

Since I've got no reply on redhat cluster list, I'm trying here ...

I'm testing a rhcs4 cluster with gfs. I noticed some poor performance while tarring data from a gfs volume to a local disk from both nodes at the same time, so I deceided to have a closer look.

Here are the bonnie++ results: http://jure.pecar.org/gfs/

I don't know what to think of these ... but I doubt it's normal for performance to go down by an order of magnitude when both nodes are doing some i/o to the shared file system ...

I'd like to see some numbers from other gfs systems to see if this is normal and if it is not, how to figure out what's wrong and how to fix it.

Thanks,