[Lustre-discuss] Lustre-discuss Digest, Vol 26, Issue 2

[Joshua Bower-Cooley]

On Monday 03 March 2008 09:00:02 am lustre-discuss-request at lists.lustre.org 
wrote:
> > What is the current "correct" way to do this now? The manual suggest not
> > using bonding, but several list postings now reccommend it. Without
> > bonding, do I need to have my 2 switches stacked, or will Lustre
> > recognize the division in my subnet?
> >
> > LNET module options I've tried are:
> > 1) networks="tcp0(eth2,eth3)"
> > 2) ip2nets="tcp(eth2,eth3); tcp(eth2) 10.9.[1-4].*; tcp(eth3)
> > 10.9.[5-8].*;" and many other variations
>
> Please use Linux bonding. Specifying multiple NICs under one tcp
> network is now a deprecated socklnd feature.
>
> Isaac

Thanks, Isaac.
I did eventually figure out the problem. Lustre was ignoring the specified 
ethernet device names and using eth0's ip to set the mgs's NID. This was 
happening with boding as well. Formatting the filesystem with only the lustre 
interfaces active solved my problem. Tunefs did not work.

Bonding seems to have performace issues with 10g ethernet. It is marginally 
slower than a single interface in every mode. No matter, our new cluster is 
all infiniband anyhow.

--Josh