Jumbo frames and NIC settings

I'm looking for a bit of help related to NIC settings on 11.2.0.3.9, 4-node cluster on RH, with jumbo frames.

We've run into a situation where an existing RAC had jumbo frames recently added. Shortly afterwards some of the instances crashed when interconnect traffic became particularly heavy. The sysadmins noted that the cards were flapping and sought changes on the cards themselves. They pulled in RedHat and RH had them increase the ring buffer size from 255 (default) to 4096 and then set tcp segmentation offload to "off". These changes helped a bit but eventually we had to turn off jumbo frames because 1 or more instances kept crashing.

The real issue is with the interconnect traffic and it's due to ridiculous parallelism settings combined with high concurrent parallel queries. Unfortunately my hands are tied at the moment on addressing this. What I'd like to do is make sure all settings across the board are intelligently set and since RH came in and suggested these changes, I started wondering if there were any best practices that others have set related to RAC and jumbo frames.

I opened an SR with Oracle asking for their recommendations (which I expect will be "mtu at 9000") but I don't have too high of expectations on this.

Dave

--
http://www.freelists.org/webpage/oracle-l