The need to tame the Internet's fractal-style traffic patterns has become all the more necessary now that circuit-switching research has virtually disappeared from the proceedings of Globecom, and asynchronous transfer mode (ATM) is rapidly fading to an afterthought. The conference's central focus on IP has brought a bigger celebrity concentration to the academic show. An IP Version 6 workshop was led by Internet pioneer Vint Cerf, and an executive panel included Hugh Martin, ONI Systems Inc. chief executive, and Larry Lang, Cisco Systems Inc. vice president of service provider marketing.

Nevertheless, IEEE Globecom remains as algorithm-heavy as it was in the 1980s and 1990s, particularly as researchers discover that most difficulties in shaping traffic at the IP layer and above fall under the mathematically tough class of problems known as NP-complete (non-deterministic polynomial-time complete). Indeed, UCLA professor Mario Gerla told one class on high-speed routing that it is unusual to find a bandwidth-shaping problem that is not NP-complete, a dilemma that is requiring the assignment of more mathematicians in the direct design of routers and Web caches.

Separate forums, common themes

This year, IEEE's Communications Society elected to break the unified Globecom conference into separate forums for broadband wireless, photonic networks, global Internet, satellite communications, communication theory, and high-speed networks, allocating the conferences among the Fairmont, Stanford Court, and Mark Hopkins hotels at San Francisco's Nob Hill. Yet except for a handful of RF modulation sessions specific to the wireless environment, the segmented conferences returned again and again to common themes of IP traffic control, a trend which Cisco's Lang said should surprise no one.

In photonic sessions, for example, several development groups advocated putting a media access control (MAC) architecture directly over wave-division multiplexing (WDM) equipment, allowing MAC-layer control of wavelengths. A research group under Maode Ma at the Hong Kong University of Science and Technology described a MAC designed for a WDM star coupler. Hajime Nakamura of KDD R&D Labs in Saitama, Japan, described a concept for "signal-free path-switching of lambdas," using a statistical lambda multiplexing topology called SLAMnet, utilizing an optical channel path assignment control mechanism.

Other photonic papers covering metropolitan rings sounded like the IEEE's recently-approved 802.17 concepts for resilient packet rings, with some added twists. Stanford University's Optical Communications Research Lab presented two separate papers on Hornet, the Hybrid Optical Ring Network, which uses an ultra-fast tunable laser to allow packet-over-WDM traffic in a two-fiber bidirectional ring. Hornet can shape bandwidth on a wavelength basis without the use of optical cross-connects or add-drop multiplexers, using only three control mechanisms-SmartDrop, SmartAdd, and Slot Manager-in its tunable-laser infrastructure.

In the RF realm, there was no end to new queuing algorithms for packet traffic in 2.5G and 3G environments such as GPRS (General Packet Radio Service) or Edge (Enhanced Data rates for Global Evolution). While the 3G Partnership Program has specified many prototype algorithms for traffic control, developers from all over the world were touting new queuing and packet-discard mechanisms to allow wireless networks to handle multiple service classes, each with different bandwidth guarantees.

The "quality-of-service" buzzwords dominated this year's Globecom to the same extent that ATM held sway over most sessions five years ago. Developers in the few remaining ATM sessions at Globecom this year reminded attendees that the QoS constraints developed for ATM are mature, and work well in public networks. But the problem of defining QoS for a connectionless service like TCP/IP has intrigued enough university and corporate research groups to make QoS a critical focus in close to half the Globecom papers.

Confronting the ubiquitous Transport Control Protocol was a subject that drew many listeners to dedicated papers at the conference. The performance of TCP has improved significantly due to the development of specialized versions in IEEE's "Nevada Cities" project, which covered three cities including Reno and Las Vegas. Nevertheless, there is much more to be accomplished in congestion avoidance. A team of developers at L.M. Ericsson's Rome Labs and the University of Rome suggested that TCP's behavior in most HTTP-based Web environments cascades into a hierarchy of fractal behaviors, suggesting that mathematical models of how TCP operates could be close to impossible to develop.

A group led by Archan Misra of Telcordia Technologies Inc. and John Baras of the University of Maryland are looking at improving TCP's response to explicit congestion notification by making the TCP congestion "window" functions more adaptive. This will require not only new software algorithms and new routing hardware, but more complex buffers in end stations and edge routers.

Improving TCP performance

Jian-Hao Hu and Kwan Yeung of the University of Hong Kong looked at how to improve Web caching in a TCP environment where wired and wireless subnetworks share access to Web servers. The research group proposed a two-layer hierarchical caching scheme called New Snoop, caching unacknowledged TCP packets at a mobile switching center and a basestation in a wireless environment. The developers suggested New Snoop could improve TCP performance in both wireline edge routers and wireless 3G networks.

Globecom even made its first foray into economics at the 2000 conference. A multinational team from the European Union's ACTS project TERA (Techno-Economic Results from ACTS) gave a sobering presentation in the aftermath of the European 3G licensing procedures, suggesting many operators will have a tough time making any money from advanced 3G services in the next few years. And a University of California at Davis research group presented a preliminary study on IP telephony, suggesting that carriers moving to an all-IP infrastructure will need to adopt totally new models of how to price voice and data calls. In contrast to the time-based calling plans in the circuit-switched world, IP networks will have to move to dynamically adjusted congestion-based pricing plans that will provide customers with better pricing choices, but might prove tough for carriers to track and bill.