Oak Ridge to develop network

Following the release in early November of the Energy Department's 20-year plan for science facilities, which stressed a big boost to the department's high-end computing facilities, Oak Ridge National Laboratory has been given $4.5 million to design a network capable of transferring huge amounts of data.

The Science UltraNet will be a dedicated circuit-switched network based on software and protocols able to handle the petascale data demanded by 21st century science, according to lab officials.

Current packet-switched networks such as the Internet are not suited to handle such massive capacities. For one thing, explained Bill Wing, a senior researcher at Oak Ridge's computer science and mathematics division, such networks have to allocate capacity according to aggregated demands from a number of different users.

"Even if you are dealing with a single stream [of data] with a carefully tuned IP stack in an uncongested network, you will still only get 5 or 6 gigabits/sec. performance," he said. "But these are heroic efforts that are not easily replicated. In such networks, bit rot sets in overnight."

Also, the way that packet-switched networks operate — by splitting data into packets sent individually over an IP network by separate routes and reassembled at their destination — introduces latencies that result in IP jitter that could wreck the close control of experiments that scientists want to have over the Science UltraNet, Wing said.

Circuit switching, on the other hand, dedicates a single circuit to a data stream until the call that began an operation is ended.

Although the Oak Ridge network will not be exceptional in basic speed, between 10 and 40 gigabits/sec., researchers will get the capability they need from optical switches and networks that dynamically reconfigure themselves and from protocols and other technologies that the lab will develop during the three-year program, Wing said.

Officials expect Science UltraNet to allow scientists to control and guide computations run on remote supercomputers, distributed collaborative visualizations and remote instrument controls.

"At its lowest level this is a networking problem," Wing said, "but sitting right next to it is the development of the middleware to enable all of this."

Brian Robinson is a freelance journalist based in Portland, Ore. He can be reached at hullite@mindspring.com.