Opening sessions confronted difficult traffic-engineering constraints involved in trying to force-fit wireless channels to the varying requirements of voice, data and video traffic. The standing-room-only QoS sessions were rivaled in size and excitement level by the one for ad hoc networks — wireless nets that must adapt to constantly varying numbers of users, without the benefit of central control.

Speakers in both tracks described a common set of problems. In QoS engineering, carriers and hardware developers must prepare for rapidly fluctuating mixes of bandwidth demands and service guarantees. Likewise in ad hoc systems, the size of logical IP subnets and the number of users can vary in ways that the developers of the original Mobile Internet Protocol never anticipated.

Rutgers University EE professor Chris Rose said it may be hopeless to attempt to map broadband traffic demands on the familiar cellular infrastructure, which is optimized for continuous coverage. Any network optimized for broadband IP delivery inherently is uneconomical for voice channels, Rose said. So IP traffic might best be handled on a separate, unlicensed band with discontinuous transmitters, like the Federal Communications Commission's 5-GHz Unlicensed-National Information Infrastructure band.

"The speed of the channel is irrelevant — if you're optimizing for IP data, the carrier can no longer make money off analog voice," Rose said.

The familiar suite of wireline distance-vector and link-state routing algorithms gets thrown out the window in a wireless world. Whether designers are building systems for single-cell wireless LANs; the discontinuous multipoint transmitters described by Rose; or cellular topologies operating in cell, microcell or picocell domains, IP packets require adaptive routing algorithms that borrow concepts from fuzzy logic and neural networks to adjust to fast-changing environments.

This is particularly evident in ad hoc network research, where routing problems are tackled by parallel or pipelined hardware, because the route-path computations they must solve are often "NP-complete" (nondeterministic polynomial-time complete) — a notoriously tough mathematical nut.

A common adaptive-routing algorithm for ad hoc networks, dynamic source routing (DSR), is used in Carnegie Mellon University's Project Monarch. Group leader David Maltz described the way the Monarch team used first-generation Lucent Technologies WaveLAN products to implement roaming DSR networks, in which the routing problem was broken in two: a route-discovery and a route-maintenance problem.

Carnegie Mellon has been using the WaveLAN system not for a fixed 802.11 wireless LAN, but for a mobile, unpredictable ad hoc outdoor metropolitan network. DSR provides a single IP address for a multinode network.

Traffic patterns

Layer 3 routing problems are relatively predictable compared with the RF constraints encountered in an ad hoc network, Maltz said: "Wireless propagation patterns are far weirder than you'd ever expect." Because the Carnegie Mellon team was operating close to a busy freeway, for example, each car that whizzed by became a source of multipath interference in the ad hoc network experiment.

An interesting contender for the DSR family of routing protocols is the ad hoc on-demand distance vector algorithm, or AODV. Elizabeth Royer, a doctoral candidate at the University of California, Santa Barbara, described a broad simulation project there for AODV protocols that's intended to take advantage of the algorithm's ability to minimize power dissipation in wireless networks by discovering IP routes only in response to source requests.

Some developers have suggested implementing AODV directly in hardware, but Royer said the disadvantage of inflexibility outweighed the performance gains of hardwiring. The Santa Barbara team implemented AODV as a user-space daemon sitting atop a Linux kernel, using a User Datagram Protocol Layer 4 socket and an Internet Group Membership Protocol multicast socket for its routing duties.

Professor Mario Gerla of the computer science department at UCLA recommended in a later speech that AODV be combined with a concept called "clustering," in order to prevent redundant transmissions in a low-power network.

Creating clusters of ad hoc nodes is a concept many have tried, but most research groups have relied on active clustering methods that require monitoring channels with high bit overhead and power consumption, Gerla said.

New work at UCLA involves passive clustering, in which nodes can passively become cluster leaders, with no active control packets transmitted in the network. The states for cluster information piggyback on the IP data packets, Gerla said, and route-state information then is transmitted by only "cluster head" and gateway nodes. The concept works well with AODV route protocols, he said, and soon will be used for DSR protocols as well.

Perhaps the best routing tactic for wireless ad hoc networks is stochastic-learning algorithms, not deterministic ones, said BBN Technologies program manager Martha Streenstrup. In a survey of routing methods for wireless networks of uncertain dimension, Streenstrup pointed to two stochastic methods: the learning-automata routing method, best used in networks with low levels of dynamic behavior; and cost-based probabilistic routing, best for networks with the highest degree of change over time.

The Defense Advanced Research Projects Agency is working with a BBN Technologies group at Verizon Inc. on a special ad hoc adaptive-subrouting layer called Dawn, for Density and Asymmetry Adaptive Wireless Network. IP packets are encapsulated for a Dawn transport delivery system that BBN has implemented in a Nokia Wireless router, using the Utilicom 2.4-GHz radio for transmission services.

Ram Ramanathan, director of BBN's internetworking research department, also described a "router in a backpack" experiment in which all equipment necessary for the Dawn transmission and routing network was tucked into users' backpacks.

To make an ad hoc IP network feasible, BBN had to develop two power-controlling packet-forwarding methods, called Lift (local information-forwarding topology) and Gift (global information-forwarding topology). Then the group worked on providing the equivalent of link-state routing through a suite of what Ramanathan called "near-sighted routing algorithms," in which the accuracy of link-state information depends in part on the hop proximity of nodes. BBN used both a nearsighted link-state and an adaptive hazy-site link-state algorithm in its network.

Packet delays studied

M. Chuah of Lucent Technologies Bell Labs provided some of the first studies of end-to-end packet delays in wireless networks. Earlier studies at Nortel Networks identified several tiers of delay for setting up a wireless packet transport, such as radio channel allocation, point-to-point protocol session setup, and transport of IP or Wireless Application Protocol packets.

But Chuah said the move to new, low-latency real-time services such as voice-over-IP will require aggregating the delays from several points in the network — the ingress and egress media gateways, the core IP network, the mobile stations, the radio access points and several other network nodes.

The International Telecommunication Union's G.114 standard recommends no more than a 150-millisecond delay for time-critical applications. This can be met easily by wireline IP networks, Chuah said, but can be achieved only occasionally in wireless nets. Delays could be increased if users turn to Session Initiation Protocol phone clients, she said, so developers should consider collapsed-network functions for wireless VoIP networks.

Some of the most intriguing QoS work for wireless IP networks is coming out of the University of Illinois in Chicago. A team under Oliver Yu is expanding on the original Internet Engineering Task Force work on Mobile IP to develop a mobile version of the Multi-Protocol Label Switching (MPLS) standard for tagging IP flows. If MPLS is seen as a "Layer 2.5" technology sitting between the media-access control and IP, Yu said, then Mobile MPLS is a "Layer 2.75" technology, using a mobile label-switched tree algorithm to map IP flows.

There are two types of reservation schemes the Illinois group proposes for wireless networks, to perform a similar function to RSVP algorithms. For controlled-load networks requiring only approximate bandwidth guarantees, the Robust Fast Reservation Protocol (RFRP) serves as a fail-safe fast RSVP. When very strict bandwidth guarantees are required, RFRP must be augmented with a second protocol, the Rerouting Synchronization Protocol (RSP), which helps minimize delay and jitter.

Yu said that Mobile MPLS methods that used both RFRP and RSP would require high degrees of hardware parallelization and pipelining, which might prove costly in silicon expense for some mobile nodes.

Yu's team will move on to work on wireless integration with optical backbones, developing a mobile version of the emerging Multi-Protocol Lambda Switching standard. The goal will be to have wireless subnets link directly to a new optical wave-division-multiplexed backbone called Star Light, which links the University of Illinois to nodes in Vancouver and New York City.

Zoran Kostic of AT&T Labs said the focus on IP alone has missed an important factor at the Layer 4 session level. The Transmission Control Protocol, often preferred over UDP because of its acknowledgement functions, could be difficult to use in traditional cellular channels. The TCP slow-start algorithms used in many TCP/IP networks, Kostic said, may face trouble when operating in networks with Rayleigh fading. And as carriers turn to the secure version of IP, IPsec, the entire packet is encrypted before sending, making it more difficult to track IP flows.

In a bandwidth-limited wireless environment, the biggest stumbling block may be the 40- to 60-byte headers of a TCP/IP packet, Kostic said. Thus, one of the most important IETF projects for wireless carriers is the Robust Header Compression working group, he said. Other packet-transport methods, such as the Enhanced Data-rate for Global Evolution (Edge), rely on stripping headers and carrying header information in a signaling channel.