A panel plots the future of data-center networks

A panel discussion at the Linley Tech Spring Conference this week tried to comb through the clamor of semiconductor marketing to tease out a future for data-center networks. During the conversation some general agreements emerged, as did some conflicts that could prove hard to resolve.

One point everyone agreed on was scale. “We will see data centers reaching unimaginable scale, with massive redundancy,” said Anshul Sadana, vice president of customer and systems engineering at Arista Networks. “We will see whole data centers backing up other data centers.”  Nor will these huge installations be confined to a few giant Web companies. “We will see enterprise data centers growing to the scale we now associate with Google,” posited John Gmuender, vice president of engineering and CTO at SonicWall.

Another point of agreement was the triumph of cloud computing, at least in concept. But there were differences about just what the concept was. On one hand there is the ideal. “We are talking about different requirements than in today’s data centers,” said Force10 Networks vice president in the office of the CTO Subi Krishnamurthy. “These data centers will be autonomous and self-provisioning.” Cisco director of marketing for the Systems Architecture and Strategy Unit Thomas Scheibe agreed. “Almost by definition, the cloud is a uniform fabric that is self-provisioning.”

Or maybe not so uniform. Sadana said he expected the links between servers to be 10 Gb, while links to the outside would move to 100 Gb. And Scheibe said he expected the distinction between LAN and WAN to endure in data centers for quite some time. It appears that the forecast collapse of inter-server, storage, and data networks into a single 10 GE mesh is not on anyone’s agenda for the near future.

The nature of the nodes on the data-center network also came in for debate. Super Micro Computer director of enterprise servers Robert Novak suggested that virtualization had allowed data-center managers not to reduce the number of server nodes, but rather to load up on very low-cost servers, in effect creating stateless computing nodes in front of very large caches. Increasingly, he said, these nodes would include GPUs. Networking and storage functions would be in separate nodes from the servers to aid virtualization. But Scheibe saw nodes moving in the other direction, toward “units of IT, each containing computing, storage, and networking resources.” Whatever the nodes look like, stateless or stateful, they need to all look the same, insisted Sadana. “To scale the network you need a uniform architecture.”

Another point of agreement was that virtualization demands security. But there was no agreement on just how to provide it. Gmuender gave a quick calculation that suggested the need for crypto-engine cycles is growing at about 100 times faster than the scale of data centers, and he recommended enhanced crypto hardware in the nodes. Novak suggested that protection for stored data might only be effective if logical files were scattered holographically across physical drives, with each stripe having a different set of keys. It did seem apparent that no one considered that one could protect an entire cloud by drawing a secure perimeter around it-encryption at the node level seems inevitable.

There was, finally, a call for unity of control. Scheibe said the data center needed a single administrator, and others pointed to the futility of trying to scale a network comprising equipment from dozens of vendors employing a Babel of allegedly standard management interfaces, many of them GUIs intended to manage one aspect of one node at a time. “A unified architecture demands standards,” Novak observed. Indeed, it appeared from the series of conversations during the panel that the entire area of data-center networks is begging for a set of standards: standards that span all the new issues that are arising as the scale of the networks increases beyond the imagination of the original network designers.