Researchers are learning that they can bring along the old mathematical tools used in ISDN and asynchronous transfer mode studies, because methodologies in queuing and traffic management apply perfectly well to routing. Papers described using fuzzy logic, simulated annealing, neural networks and even wavelet-based digital filtering to manage and prioritize traffic in order to carry real-time services on IP routers.

The emphasis on routing stood in sharp contrast to the Globecoms of a decade ago, when only a handful of papers touched on IP routing. "Part of the problem was that 10 years ago, Cisco and Wellfleet were doing all the routing research in secret, at their own skunk works," said one Stanford undergraduate who did not want to give his name. "But it's also true that at the universities, IP is now cool. It was always cool to the Internet weenies in academia, but now it's cool to the hard-core telecom researchers, and that wasn't always true."

Now that multiprotocol label switching and its optically enhanced cousin, Generalized MPLS, have been universally accepted as the flow-management model of choice, researchers want to play with it. Shai Benjamin and Izhak Rubin headed up a program at the University of California, Los Angeles, to develop MPLS+, using label swapping in conjunction with path merging and intermediate-destination removal. The idea is to minimize the size of switching tables, thereby simplifying MPLS further.

A team at Syracuse University wants to simplify the creation of label-switched paths for MPLS inside a routing domain by means of a new network element, a bandwidth-management point, responsible for calculating service demands and network resources. A group under Kenji Sakamoto at Keio University in Yokohama, Japan, meanwhile, has designed a special virtual-circuit merge switch, so that ATM routers that use MPLS do not have enormously big buffers to handle cell reassembly functions.

One group at the University of Texas at Dallas moved even closer to silicon, designing a dedicated search engine based on content-addressable memory, optimized for MPLS over ATM. The million-transistor chip, fabricated by Canadian Microelectronics Corp., diverged from standard binary or ternary CAMs, since it implemented a special cache-based sorting algorithm developed by Yiyan Tang and Yuke Wang at UT.

IBM's Zurich Research Lab is bypassing CAMs for handling the very large routing tables that are expected to become the norm as Internet Protocol version 6 replaces IPv4. Jan van Lunteren, a researcher at IBM Zurich, described a new IP lookup scheme called Barts, or balanced routing-table search, based in part on a concept of encoding a prefix tree to create a table that looks like a hashing function.

Combinatorial logic is used to create a counter for a compressed table, and the search logic then can be integrated on a chip with a block of embedded memory, such as SRAM. Van Lunteren said early simulations showed that wire-speed route table lookups could be performed at 10 Gbits/second using IPv6 addresses.

While many quality-of-service (QoS) sessions dealt with the algorithms to implement in coprocessing silicon for routers and switches, true network processor papers were rare. H. Michael Ji of Tensilica Inc. prepared a pair of papers on the use of reconfigurable processors in packet-classifier problems, but Tensilica was the only silicon vendor giving public implementation examples.

Many designers from academia appear sensitive to the commercial world's need to keep solutions both simple and backward-compatible. Sudeept Bhatnagar and Brett Vickers of the Rutgers University computer science department described a special protocol, Edge-assisted QoS, intended for networks where only the edge routers can participate in QoS tagging. And Raquel Hill and H.T. Kung of Harvard are aiming at a collapsed Differentiated Services and call-admission control function, offering a new protocol, DEAC, that sets up a gatekeeper for call admission and establishes priority classes for IP traffic.

Developers continue to ascend the protocol stack to the session layer, looking for enhancements or replacements for the venerable Transmission Control Protocol to better handle mixed-traffic sessions. Each year, at least one derivative of TCP named for a city (TCP Reno, TCP Tahoe, TCP Vegas) gets its debut at Globecom, and 2001 was no exception.

A UCLA team presented two papers on TCP Westwood, which maintains compatibility with TCP Reno but improves congestion control by implementing a measured way of shrinking the session "window" when there is packet congestion. TCP Reno divided the window size in half, which turned out to be too inefficient.