On 2/4/2010 11:59 PM, Carsten Aulbert wrote:
(1) ping -s 8972 -Mdo<remote host> (try different payload sizes and remember that there might be some overhead in the switches needed)
This results in icmp_seq=2 Frag needed and DF set (mtu = 1500) which is not what I expected. I wonder where the "mtu = 1500" is coming from. ifconfig on the interface of the source machine is definitely 9000 (I just reconfirmed). I also confirmed that jumbo frames are enabled both on the switch and on the storage server. They are. For yuks, I tried lowering the packet size lower, and I found that I continued to see this error until the packet size was 1472. So, either ping is doing something I don't expect, or somebody is lying about jumbo frames being enabled.
(2) Use netperf between different nodes and see if the performance is not drastically reduced with large jumbo frames.
The funny thing is that performance seems to be fine, although that's purely subjective. I'll try the netperf test to see what the numbers really are.
Would it be possible for you to recompile the kernel with the same settings and enable timings in printk lines (under kernel hacking)? That might help, but might be some work to get working.
This is a 48 node cluster, so doing something like that is something I'd like to hold off on doing until I've exhausted everything else. I appreciate your suggestions. -- Jon Forrest Research Computing Support College of Chemistry 173 Tan Hall University of California Berkeley Berkeley, CA 94720-1460 510-643-1032 [email protected] _______________________________________________ autofs mailing list [email protected] http://linux.kernel.org/mailman/listinfo/autofs
