By Siobhan Kennedy
A spokesperson for Intel Corp said this week that its NGIO server specification could work just as well over internet protocol as the rival Future I/O architecture, despite claims made to the contrary. Earlier this week, networking giant Cisco Systems Inc said it was backing the Future I/O spec, from Compaq, IBM and Hewlett-Packard, over Intel’s NGIO. Speaking to ComputerWire, Frank Maly, director of marketing for Cisco’s interworks business division, said the company made the decision to go with Future I/O because of its native support for IP. Having IPv6 incorporated in the spec will enable customers to tightly couple their I/O based server-to-server, and server-to-storage networks with their data networking infrastructure, he said. Today companies typically have to keep these networks separate, which is more expensive, but combining the two will enable businesses to use their internal intranet, or the internet, to carry out such tasks as back-up and disaster recovery, he said.
Mitch Shultz, Intel’s director of server platform marketing, admitted that NGIO isn’t based on IP but he denied that that meant the I/O traffic couldn’t be transported over an IP-based network, like the internet. That’s simply wrong, and it represents a serious misunderstanding of what NGIO is all about, he said. Shultz said Intel had deliberately not gone to all the effort of incorporating native IP within the spec because, on balance, it decided it was not necessary given the nature of I/O traffic. We looked at it but we concluded that, for the type of traffic, the overhead of utilizing IP far outweighed the benefits, he said, There’s no benefit to carrying native IP data structures inside a system area network.
But that doesn’t mean the traffic won’t run on an IP network, Shultz added. It just means that when it gets to the router, the I/O data will need to have instructions added on top of it, like a tag, to ensure it gets to the right destination. That’s exactly what Cisco does itself at the moment with its tag switching, he said, and all references to the nodes are based on the tags, not the IP packet. I just don’t understand what they’re saying.
But Cisco’s Maly yesterday stood by his statement that not fully integrating with IP was the main reason the networking giant was not giving its full support to the NGIO specification. It’s a complexity issue, he said, yes, you can still send NGIO traffic over IP networks but first you need to manually define the end points of all the connections and that’s a complicated procedure. Maly said to infer you could just put an IP tag on the I/O data was trivializing the issue. He said the fact that the Future I/O group was trying to integrate IP into its spec from day one was key. If you carry it [IP] natively, it’s a much simpler procedure and a much cheaper solution, and it’s easier to manage for the customer.
