I've been following Infiniband developments for many years now and have myself baffled by the cultural resistance to this technology. There seems to be an overwhelming feeling that ethernet will evolve, or fibre channel will suffice, or that somehow we won't need more bandwidth that we have!

I remember an ex-professor of mine claiming that fortran would never die and there was no need for any other programming language. As he put it "I don't know what the number one programming platform will look like in 10 years time, but it will be called fortran"! ... it may well be an applicable attitude here - the facts behind our networking demands and related challenges speak for themselves, when we will finally recognise that we (the industry) need a new approach and embrace infiniband as the best we have to work with?

The oddest thing about the "ethernet will take over" positing is the comparative pricing of infiniband to 10GE. Infiniband is much cheaper. You'd think that it would therefore be "winning'.

Of course, if you look under the hood at the deployment issues, I'd say that if 10GE were here today at a good price point, and you were buying the right cards, you'd want to go with that. There are a few 10GE cards out there that can do RDMA, and therefore occupy the Infiniband niche. Conversely, when considering the use of Infiniband to occupy the ethernet niche.... well who wants to fuddle with IP-over-Infiniband drivers and the like?

Joe Kraska

San Diego CA

USA

Just realised I said "need more bandwidth" above - what I should have said was "need more effective networks" ... reason being it is more likely that latency with be a defining factor.

I would also pick up on Joe's comments above - which are spot on by the way - when we are looking at 10Gb/s plus via ethernet we are going to have to look at TOEs and OS enhancement to optimise (or circumvent) the stack. Having vendors bundles IP-over-infiniband drivers shouldn't be too painful ...

... then we can also put the energy efficiency factors into the mix, as well as considering what we do when demands extend to 100Gb/s and above.

I think most pressing arguments revolve around the opportunities architectures such as inifiniband provides for looking at out infrastructure architectures with fresh eyes and exploiting "new" (are there any really new ones recently?) approaches, such are a true grid design for example.

It's interesting that the article can't tell the difference between gigabit (Gb) and gigabyte (GB)!! After all, one is only 8 times more than the other. A mere trifle.