1
Enabling Large-scale Wireless Broadband:
The Case for TAPs

R. Karrer, A. Sabharwal, E. Knightly
Rice University, Houston, TX

{karrer,ashu,knightly}@ece.rice.edu

Abstract--- The vision is tantalizing: a high-performance, scalable,
and widely deployed wireless Internet that facilitates services
ranging from radically new and unforeseen applications
to true wireless "broadband" to residences and public spaces
at rates of 10s of Mb/sec. However, while high-speed wireless
access is easy to achieve in an enterprise network via low-cost
IEEE 802.11 (WiFi) access points, wireless technology in public
spaces is in its infancy. "Hot spots" provide high-speed wireless
access, but do so in very few isolated "islands" at immense costs.
Likewise, while fixed wireless (e.g. LMDS) and 3G can provide
ubiquitous coverage and 3G can support mobility, throughputs
can often be two orders of magnitude slower than WiFi.
In this paper, we formulate the challenges of building a highperformance,
scalable and widely deployed wireless Internet
along 10 premises. We make the case for the requirement of a
fundamental new architecture based on beamforming antennas
deployed on fixed, wire-powered Transit Access Points (TAPs)

that form a multi-hopping wireless backbone with a limited
number of wired ingress/egress points. To address scalability,
deployability, and performance challenges we present distributed,
opportunistic and coordinated resource management problems
and a novel "network is the channel" framework that searches
for fundamental information-theoretic tradeoffs between protocol
overhead and capacity.

I. INTRODUCTION
Over the last decade, we have witnessed an explosion in
wireless access to the Internet. In 2002, revenue from IEEE
802.11 (WiFi) network cards and access points totaled an
estimated $2.1 billion on 23.9 million devices with 73%
growth predicted for 2003 alone.

1

Moreover, advances in
the physical layer and media access protocols have enabled
transmission rates of 54 Mb/sec in IEEE 802.11a, and even
higher rates are projected in future revisions.
However, in spite of these advances, we remain in the infancy
of of the long-standing vision of a high-speed ubiquitous
wireless web. To date, the overwhelming majority of deployed
WiFi networks are in the enterprise or home, restricting highspeed
wireless data communication to small wireless "islands."
There are two simultaneous efforts to providing wireless
Internet beyond these islands. The first is deployment of
"WiFi hot spots", typically consisting of IEEE 802.11b access
points connected to a (T1) wired backbone. However, the
great fanfare with which each hot spot is announced [1] is
immediately tuned down by sheer numbers: at the end of year
2002, the U.S. had approximately 3,000 hot spots, attracting
an estimated 20,000 users, resulting in a net revenue of $2

This work was partially supported by National Science Foundation under
Grant ANI-0325971.The authors will like to thank their collaborators,
Behnaam Aazhang, Patrick Frantz and Dave Johnson for invaluable discussions
and comments.

1

Source: Gartner Dataquest.

million -- and yielding a large net loss, given fixed costs as
high as $10k and recurring costs of approximately $400/month
per hot spot.

2

Ironically, the overwhelming costs of providing
wireless hot spots is due to fixed and recurring costs of the

wired infrastructure.
As a consequence, deployment is low and "coverage" is
dismal. Even adding Cometa's plans for an additional 20,000
hot spots to the existing 3000, and optimistically estimating
that each hot spot covers 100x100 m

2

, coverage will be as low
as approximately 4 km

2

per metro area, or 0.4% of the area
of a moderate sized city such as Indianapolis. Thus, today's
hot spot architecture is slow to deploy, costly, and unscalable,
and is not on any path to provide large-scale coverage.
A second major effort is 3G and fixed wireless services such
as LMDS. However, in both cases, speeds are typically 2 to
3 orders of magnitude slower than WiFi, with maximum peruser
speeds in the 100s of kb/sec range. Moreover, because
of multi-billion dollar spectral license costs and high infrastructure
costs, such systems have proven costly to deploy and
hence lead to expensive, yet moderate speed, wireless Internet
services. Thus, while having the promise of near-ubiquitous
coverage and allowing high mobility speeds, such technologies
have significant performance and cost limitations. Moreover,
given their small Internet subscriber base, scalability to many
data users remains unproven. Therefore, despite a decade of
strong progress in wireless data communication, it is clear that
with the current evolutionary path, a large-scale high-speed
wireless web is not on the horizon.
This paper describes the challenges of building a wireless
Internet that simultaneously achieves deployability, scalability,
high-performance, and a cost-effective economic model along
10 fundamental premises. We believe that these premises -
some of which are inherent to any wireless network, some
of which are specific to the outlined challenges - provide the
basic framework to realize the above vision.
II. THE CASE FOR TAPS

Premise 1: Designing a wireless Internet that simultaneously
achieves deployability, scalability, high-performance,
and a cost-effective economic model requires a new architecture.
This architecture is based on fully-wireless
beamforming Transit Access Points (TAPs) that form
a multihop backbone mesh which interconnects TAPs,
mobile units (MUs), and the wired Internet.

The dominant infrastructural costs of traditional hot spots
lead to the (logical) conclusion that many of the AP's wires

2

Sources: Jupiter Media Metrix analyst Dylan Brooks and Insight on
Wireless analyst Andrew Luan.

2
H
G
F

Internet
Internet

A
B
C
D
E
I
Directional Transmission
Fig. 1. TAP network
must be removed. However, simply removing the wires creates
an ad hoc-like network where packet forwarding of wireless
APs and user access compete for the scarce spectrum, pushing
the system capacity dramatically to the well-known scalability
limits of ad hoc networks [9].
We envision therefore an architecture as depicted in Figure
1, where TAPs equipped with sector antennas allow for
geographically focused as well as omni-directional transmission
of data. Being equipped with multiple antenna arrays,
TAPs can form multiple orthogonal beams and hence communicate
with different destinations simultaneously. Likewise,
simultaneous transmissions can be separated in frequency via
use of multiple orthogonal subbands (e.g., IEEE 802.11b
has 3 non-overlapping channels and 11 channels within the
WiFi unlicensed spectrum). In practice, the actual number of
possible simultaneous transmissions will be limited by the
number of actual air interfaces that can be mounted on a TAP.
Since TAPs are not mobile, their relative spatial location
does not change. This stability allows the use of directional
transmission, known as beamforming. Beamforming improves
the system throughput in two ways. First, there is an increased
received energy at the destination as well as a higher per-link
capacity because beamforming does not spread its energy in
all directions. Second, directional transmission creates little or
no interference to ongoing transmissions to and from mobile
units, which increases spatial reuse.
Building a TAP architecture introduces new research challenges
at the physical layer. First, state-of-the-art beamforming
techniques [4] assume that only either sender or receiver
are equipped with multiple antenna elements, but the TAP
architecture assumes both. Second, MIMO space-time encoding
(e.g., [26]) assumes that antenna elements are spaced
sufficiently far apart to create independent fading at each
element so that the antenna beam patterns are not focused,
whereas TAPs require focusing.
Thus, using directional antennas, an interconnected TAP
wireless "backbone" can be formed with high speed and a high
degree of spatial reuse. This backbone efficiently forwards
traffic from and to multiple wired TAPs, which additionally
have a connection to the wired Internet with possible capacities
up to 100s of Mb/sec (e.g., Ethernet, Gigabit Ethernet, and
OC-X access links). Since the TAPs may not necessarily
provide complete coverage for economic or environmental
reasons (obstructions), mobile users, such as G and H in
Figure 1, can have their packets forwarded by other mobile
users over multiple hops before reaching a TAP.
This combination requires fundamental research in deriving
the transmit and receiver array coefficients to maximize signal
to interference plus noise ratio (SINR) at the receiver while
ensuring that the ongoing transmissions do not suffer any
degradation in SINR.
III. COORDINATED AND OPPORTUNISTIC RESOURCE
MANAGEMENT

To achieve system-wide high performance, the TAP network
must address fundamental new challenges to coordinate and
opportunistically exploit available resources system-wide.

Premise 2: Opportunistic selection of high-quality paths,
sub-bands, and channels is required due to fast timescale
of variations in channel conditions and the availability of
multiple paths to and from wires.

The traffic behavior of a TAP network is unique in two
ways. First, unlike cellular and ad hoc networks, traffic does
not have a unique, fixed destination, but rather can be delivered
to "any wire." In Figure 1, data from MU I can reach the
wire via TAP B or E. Based on prior information about the
end-to-end available bandwidth on each route (Section IV)
and fast timescale channel measurements (Section II) traffic
can opportunistically be scheduled to the best current path.
Second, TAPs are equipped with more than one air interface,
thereby enabling more than one simultaneous channel via
beamforming or orthogonal frequency bands.
The scheduling challenge is to design a distributed opportunistic
multi-channel, multi-destination scheduler. In an
ideal case, a scheduler could utilize information regarding
channels and queue backlogs of all flows to maximize throughput
by exploiting high-quality channels, best-quality paths to
different wires, and multiple air interfaces. In practice, this
decision must be made while incorporating the distributed
nature of TAP resource coordination (TAPs do not have perfect
knowledge of other TAPs and MUs), subject to constraints on
limits on the number of simultaneous transmissions (imposed
by the number of air interfaces), and subject to balancing
transmission of ingressing and transit traffic to provide a fair
allocation of time shares in the system.

Premise 3: Avoiding contention by adaptively selecting
backoff times allows the system to scale.

Scalability is seldom associated with scheduling and
medium access protocols. Yet, as the number of users increases,
the amount of side information (about channels and
queues) and overhead in contention resolution increases without
bound for current protocols, making them unfit to scale
for our envisioned system.
Common random access MACs are limited in scalability
because contention incurs long backoff periods and high
collision rates, thereby severely throttling system goodput
and increasing delay. If a contention-free system is assumed,
information theoretic bounds predict that as the number of
users in the system increases, the net throughput should increase
unboundedly [13]. Thus, it is not evident from existing
results whether non-scalability of contention resolution MACs
is fundamentally unavoidable.
To facilitate design of load-scalable medium access protocols,
the average time spent in contention per packet must be

3
Fig. 2. Coordinated resource management
traded with queuing delay. For example, if a node acquires
the channel and retains it for a duration of multiple packet
transmission times, then the time used in contention per packet
is effectively reduced by a corresponding factor. To achieve
scalability, the time that a node retains the channel should
increase with the number of contending flows. Thus, while
the delay in this system necessarily increases with the number
of nodes, the system goodput can scale. By further scaling
the number of consecutive packet transmissions in direct
proportion to the current channel conditions, MAC scaling
can be integrated with opportunistic scheduling while still
maintaining compliance with IEEE 802.11.

Premise 4: Any medium access and scheduling decision
requires distributed resource management rather than a
purely local decision.

Unlike schedulers designed for cellular and wired-AP networks,
(e.g., [3], [5], [11], [17]), scheduling in TAP networks
is inherently a distributed operation. Nodes in the network
are not aware of the channel conditions or queue backlog of
other nodes. It is evident that naive exchange about everyone's
local information will lead to protocol overhead explosion
and in turn, a scheduling discipline that will not scale. Thus,
new techniques are needed to enable a scalable opportunistic
scheduler for networks with distributed control.
In a first step, a centralized solution for opportunistic
scheduling may provide an upper bound to performance. In
a second step, the broadcast nature of the wireless medium to
share information can be used to ensure scalability. Namely,
by piggybacking information on data and control packets such
that other nodes can overhear, nodes can obtain a partial,
but necessarily incomplete view of the "distributed queue."
Thus our thesis is that with perfect information and centralization,
the net system throughput will grow with an increasing
number of users even under fairness constraints (as in [13]),
whereas with partial information sharing, scalability can be
maintained albeit with stochastically bounded deviation from
the centralized solution. Note that in contrast to centralized
systems where scheduling and medium access are typically
addressed independently, the distributed nature here implies
that their effects are tightly coupled, adding to the challenge
of this problem.

Premise 5: Coordinated resource management is required
to eliminate spatial bias of throughput and to exploit
spatial reuse.

The TAP network must ensure that all nodes in the network
receive a proportionately fair share of the network capacity.
Our view of fairness is that a node should get the same
bandwidth share independent of whether the node is just 1
hop away from a wired TAP or whether reaches the wire via
multihopping. Consider the scenario of Figure 2. Suppose that
the link capacity to the wired Internet is the current bottleneck.
If the TAPs provide only local fairness, each of the depicted
nodes communicating with wired TAP 3 (MUs 7 to 10 and
TAP 2) would receive an equal bandwidth share to the wire.
However, wireless TAP 2 requires a far greater bandwidth
share than MUs 7-10, as it is forwarding aggregated traffic
both from its own serviced mobiles (MUs 2-6) as well as
aggregated traffic from farther upstream (from TAP 1).
Consequently, flows must be throttled to ensure fairness.
This throttling must be done at the first TAP to achieve
efficient spatial reuse, and therefore scalability. Returning to
the scenario of Figure 2, suppose that the flow from MU 1 is
bottlenecked at the wire of "Wired TAP 3" to a "fair" rate of 1
Mb/sec and that MUs A communicate only locally. Then, only
by throttling the flow of MU 1 by TAP 1 can MUs A and B
use the full remaining capacity for their local communication.
This flow throttling is explicitly necessary for a TAP network,
although, at a high level of abstraction, TCP addresses
fairness and spatial reuse via additive-increase multiplicativedecrease
congestion control. However, relying on TCP alone
is not enough. First, TCP's congestion control has welldocumented
performance limitations over both multi-hop and
single-hop wireless networks (e.g., [2], [6], [10]). Second,
TCP's congestion control necessarily operates at end-to-end

timescales of 100s of milliseconds -- too coarse to address the
fast timescale dynamics of contention and realistic channels.
Finally, TCP naturally biases flow throughput to favor flows
traversing fewer hops. In contrast, the objective of a TAP
network is to provide fair or minimum bandwidth targets
independent of spatial location.
Likewise, significant progress has been made in distributed
media access and scheduling algorithms designed to balance
fairness and spatial reuse objectives in ad hoc networks (e.g.,
[18], [22], [27]). There are two critical aspects of the TAP
network that require a fundamentally new look at distributed
resource allocation. First, the network has a distinct structure
as compared to general ad hoc networks because the TAPs act
as points of centralization through which most traffic passes.
Namely, combined with the use of directional antennas, the
TAP network has a unique concept of transit traffic traversing
a backbone. Second, as described below, the performance
objective (fairness reference model) is different for TAPs as
compared to general ad hoc networks.

Premise 6. `TAP-aggregates", and not MU flows, should
be the basic fairness element.

Returning to the example of Figure 2, our notion of fairness
is that all TAPs should get the same fair bandwidth share. As a
consequence, however, not all MUs are given the same share:
since TAP 2 is serving more MUs than TAP 1, MUs 2-6 are
given a smaller share than MU 1.
Since it is impossible to achieve TAP-aggregated and perMU
fairness, we advocate for TAP-aggregated fairness for
three reasons. First, TAP-aggregated fairness provides exactly
the same service level that would be achieved if each of the
TAPs were a traditional "wired" hot spot, namely, the MUs

4
equally share the capacity of the local wireless channel and
the wired link. Second, a fairness reference model of TAP
aggregates enables us to design scalable coordinated resource
management algorithms that would not be possible with perMU
approaches. Finally, TAP-aggregated fairness removes
spatial bias of throughput that would occur with only local
fairness mechanisms.
Spatial bias of throughput must be addressed by designing a
formal reference model for achieving fairness and spatial reuse
in TAP networks. This TAP-aggregated fairness model differs
fundamentally from both proportional fairness as approximately
achieved by TCP [12], [19], [20] and max-min fairness
as targeted by some ATM congestion control algorithms [14].
While the solution to achieve this desired reference model
for the scenario of Figure 2 is immediate, the general case
provides significant challenges due to variable rate channels,
MU mobility, dense TAP meshes, bi-directional traffic, etc.

Premise 7. New coordinated resource management algorithms
are required to achieve the TAP-aggregated fairness
reference model.

For the above reasons, the reference model must provide
a coordinated and distributed resource management algorithm
and protocols that have both a proactive and reactive component.

The proactive aspect of such protocols must consist of
messages exchanged among TAPs to convey information about
a TAP's aggregated traffic demand and channel conditions.
With this information, TAPs can make a coordinated decision
as to the relative service rate of ingressing and transiting
traffic. The objective is to balance throttling flows to their
bottleneck fair rate with more aggressive forwarding that
ensures that a sufficient number of packets are backlogged
at TAPs to exploit opportunistic medium access when high
quality channels permit, or when contention and congestion
is temporarily reduced. Addressing this issue requires the
development of a performance analysis framework to gain
fundamental understanding of the relationship between local
channel-dependent medium access decisions and system-wide
performance.
The reactive aspect must operate on a per-packet basis (versus
per-TAP and per-MU throttling). Here, the critical issue
is to ensure that each packet meets its targeted performance
objective, despite multi-hopping across highly variable channel
conditions. To solve this problem, coordinating packets' priority
indexes among nodes is essential. In wired networks, a
class of coordinated schedulers has been developed that allows
packets that are "late" or under-serviced upstream to catch
up at downstream nodes by coordinating a packet's priority
index across multiple nodes [15]. In TAP networks, multihop
coordination to best achieve system-wide performance
objectives must take variable channel conditions into account
and must interact with the random access MAC protocol.
IV. THE NETWORK IS THE CHANNEL -- ESTIMATION,
PROTOCOLS AND CAPACITY SCALING
In addition to the protocol design, which is driven by
capacity and scaling issues, the TAP architecture also provides
a unique possibility for protocol-driven capacity analysis.

A

(a) (b)

B
C
D
E

F

G
Internet

C
D
I
H
B
A
Source
MUs
F
G
E
Wireless TAPs
Destinations
H I
Internet
Fig. 3. Network channel depiction for the wireless TAP architecture for MU
to wired TAP communication.
Premise 8: Information-theoretic channel capacity analyses
are overly optimistic because they ignore the performance
impact of protocols (e.g., MAC, scheduling, and routing). A
new view of the "whole network is the channel" is needed
to understand fundamental tradeoffs between protocol
performance and system capacity.

The key to a high-performance scalable system is to ensure
that packets consume minimal system resources to reach
their destination. In particular, the scalability limitations of
purely ad hoc networks [9] arise because each forwarding
hop consumes additional resources. Moreover, equally crucial
scaling impediments can be observed in measurement studies
[8], [16] which show that actual implementations perform
significantly worse than the predicted information theoretic
bounds [7], [9] because they assume perfect "zero-overhead"
protocols. Thus, while representing an important step in understanding
the behavior of large-scale wireless ad hoc networks,
existing theoretical capacity results provide limited insights on
system design issues. But to understand the real-world scaling
behavior of the TAP architecture as well as ad hoc networks
in general, a capacity analysis that incorporates the critical
impact of protocols is essential.
The design of a routing protocol for the TAP network, e.g.,
must contend with two unique issues not previously addressed.
To achieve high performance, it is essential that the routing
protocol consistently discover high-quality routes. However,
this discovery must be balanced with the resulting routing
protocol overhead. Second, the TAP network is inherently heterogeneous
in terms of power limitations, transmission ranges,
channel qualities and (wired and wireless) bandwidth. Thus,
routing protocols must contend with a dynamic and highly
non-homogeneous TAP backbone in addition to mobility and
dynamics encountered in ad hoc networks. The challenge is to
develop an analysis and protocol design methodology based
on treating the whole network as a channel, which clearly
identifies the role and the impact of protocols.

Premise 9: The "network is the channel" framework allows
for an integral solution that addresses the heterogeneity in
timescales and transmission modes in a TAP.

The spatial distribution of MUs and TAPs, as depicted in
Figure 3(a) is noted by an individual MU as a composite channel
between itself and its destination (in most cases the wired
Internet). This notion of a composite channel, labeled network
channel, is depicted in Figure 3(b), with MUs, wireless TAPs
and wired TAPs represented in different sets to emphasize
their difference in power limitations and capacity. Analogous

5
to any other channel studied in information theory, the network
channel has a capacity.

3

To understand the fundamental limits
on protocol overhead and the network channel capacity, it
is necessary to study the different timescale variations and
transmission modes.
The fastest timescale variations (on the order of several
packets) impact the performance of beamforming and opportunistic
scheduling which utilize channel measurements made
at that timescale. A fundamental bound on the capacity of
beamforming for a system with M transmit antenna elements
and a single receive antenna using B bits of channel information
was presented in [21]. First of its kind, this bound
uses no asymptotic approximations and is thus valid for all
practical cases of interest. These results can form the basis to
study the relationship between channel coherence time and the
channel measurement rate for TAP to TAP and TAP to MU
communication, where the receiver can have more than one
receive antenna element.
At longer timescales, variations in traffic patterns, channel
conditions, and contention impact network capacity such that
coordinated resource management using message passing is
essential for fairly throttling flows and maximizing throughput
and spatial reuse. While increased protocol information on
network channels at this time scale can provide increasingly
precise control, the overhead of message exchange will
eventually overwhelm performance. A delay-limited capacity
theorem characterizes the fundamental relationship between
queuing delay and average transmit power for single link communication
[24]. Based on this work, new capacity results can
be derived that consider the case that only limited information
is available to MUs from fast timescale channel estimation and
coordinated resource management thus providing a realistic
characterization of capacity and scaling.
Finally, at the longest timescales, node mobility leads to
unpredictable changes in the probability distribution function
(pdf) governing channel variations. This leads to a fundamentally
different situation compared to traditional information
theoretic analysis where transmitters and receivers are assumed
to know the channel pdf. However, in multi-hop networks like
TAP networks, nodes are unaware of the network channel pdfs
and must estimate them as a precursor to actual communication.
With the above conceptual organization, we observe
that routing protocols are network channel estimators.

4

Similar
to the establishment of the relationship between the number
of channel measurements and long-term route throughput for
simple linear topologies [25], the routing protocol overhead
is related to the level of network mobility and the resulting
system capacity.
These analytical tools provide critical foundations for a
complete scaling analysis that incorporates protocol overhead
in measuring fast and slow timescale channel variations, the
impact of traffic patterns on spatial reuse, and the relationship

3

The capacity of the network channel is the maximum rate at which the
source node can transmit such that it can be reliably (with vanishingly small
probability of error) received at the destination.

4

Note that the objective is not necessarily to form a highly accurate estimate
of the network channel, but rather to obtain an estimate that satisfies the
routing objective such as finding a minimum hop path subject to performance
constraints.

between routing overhead, mobility and quality of discovered
paths. Such an analysis is particularly crucial for a TAP
network: because it includes protocol overhead at various
layers, it can already be employed in protocol design to study
scalability and throughput limitations.

Premise 10: The "network is a channel" view allows for
designs of hybrid and scalable routing protocols.

Since the bandwidth and stability of MU and TAP links
differ significantly, a two-tiered hybrid routing protocol is
required to exploit node heterogeneity. In particular, because
the TAP to TAP links have relatively high reliability and
bandwidth, a proactive (periodic) routing protocol is needed
for TAP to TAP routing. In contrast, routing to and from MUs
can be reactive to address mobility and the variable channel
dynamics of MUs.
To address the scalability challenge, the following two key
innovations are required. First, by decoupling the TAP to TAP
routing from routing involving MUs, DSR-like routing only to
the first TAP ensures that requests originating from MUs never
traverse the TAP network. Furthermore, the route request from
MUs can target the nearest TAP(s) and can be restricted to
traverse the MU's local neighborhood. This restriction bounds
the average path length traversed by route requests, resulting
in improved traffic scalability [16], and providing a foundation
for scalable routing. Moreover, by exploiting the network is
the channel framework, the overhead in discovering new and
better routes can be balanced with the quality of the resulting
paths.
Second, a scalable location management protocol is required.
We envision a distributed system of "home agents",
similar to Mobile IP [23], located at TAPs. In particular,
each mobile unit can register with the closest home agent.
This agent is also used to discover the intended MU for
traffic originating from a wire. The registration can be reactive

and be performed in the process of uplink route discovery
initiated by MUs. Note that this contrasts to the current cellular
approach, which is completely proactive in nature. During
route discovery over the TAP backbone (which is not the same
as MU to TAP route discovery), either one or more TAPs can
be associated with the intended MU.
V. INDUSTRIAL EFFORTS

The long-standing vision of a high-speed ubiquitous wireless
web has also attracted several companies. Ricochet Networks,


5

a daughter company of Metricom, was the first to
deploy a commercial architecture with multi-hop wireless
transmission consisting of a grid of proprietary "radio receivers
" spaced within a half mile of each other, and covering
17 metropolitan areas. Unfortunately, Metricom's approach led
to economic failure and eventually bankruptcy in July 2001.
While quite innovative compared to alternative solutions at
the time, Metricom failed technically at many levels in both
its architecture and protocols. It did not achieve scalability
(the 50,000 subscribers were spread over half as many radio
receivers), nor high performance (peak rates were limited to
128 kb/sec), nor cost-effective deployability (high deployment

5

http://www.ricochet.com

6
and operating costs without exploitation of economies of scale
for many users resulted in high subscription costs and a small
subscriber base). Such past failures highlight the need to
rethink the fundamentals of algorithms and architectures for
large-scale wireless systems and illustrates the requirement to
leverage the attractive economics and installed base of existing
IEEE 802.11 hardware.
There are also many ambitious industry efforts that provide
a small piece of the solution for a wireless Internet, such as
directional antennas (e.g., AirNet, SkyPilot, Vivato), multihopping
(e.g., MeshNetworks, RoofTop Communications), IPcentric
base stations (e.g., Flarion), and Hot Spot operators
(e.g., Boingo, Cometa, T-Mobile). While their success or
failure is not yet established, the missing link for achieving
scalability, deployability, and high performance is not simply
integration of these components, but rather requires holistic
and fundamental research into the foundations of TAP-like
architectures, an objective that is not being addressed by any
current industrial effort.
VI. CONCLUDING REMARKS

The development of the described TAP architecture impacts
a set of critical application scenarios. First, by removing the
dominant costs of hot spots associated with wired infrastructure,
a wireless TAP network will provide an economically
viable and deployable architecture to provide large-scale high
speed wireless access to large user populations. In particular,
TAPs will exploit the cost-effectiveness of mass market
wireless devices that have driven markets to the $50 access
point, a cost-performance curve that cannot be achieved by a
fully wired AP infrastructure due to the physical necessities
of wires (such as the expenditures of digging trenches and
laying cables) and their device requirements (such as router
line cards). These advantages will enable large-scale WiFibased
deployments with broad coverage versus today's smallscale
hot spot islands.
Second, TAPs provide a key technology for true broadband
to the home. Today's broadband-to-the-home efforts require
that each person independently purchase a relatively low speed
(100's of kb/sec) "broadband" connection from an ISP. The
resulting high costs, moderate data rates, and requirement to
use existing infrastructure (phone or CATV lines) has resulted
in disappointing service and a dismal penetration rate of less
than 10% of households. With TAPs, communities (through
local governments) or new access providers can purchase a
neighborhood T3 connection (for example) and deploy poletop
TAPs to provide low-cost, high-performance broadband to
the home. Since the TAP network aims at maintaining IEEE
802.11 compliance, users are not required to buy expensive
cards. They rather can use the same WiFi cards at the office
and wherever a TAP network is available.
Thus, by addressing the outlined challenges, the TAP network
provides a critical foundation for the wireless Internet,
and has the potential to transform from today's frustratingly
slow, overpriced, unreliable wireless data services into a new
wireless Internet at an unprecedented scale, economy, deployment,
and performance.
REFERENCES

[1] A. Baer. The Wi-Fi boom; out and about, and online. The New York
Times, December 12, 2002.
[2] H. Balakrishnan, V. Padmanabhan, S. Seshan, and R. Katz. A comparison
of mechanisms for improving TCP performance over wireless links.
In Proceedings of ACM SIGCOMM'96, Stanford, CA, August 1996.
[3] V. Bharghavan, S. Lu, and T. Nandagopal. Fair scheduling in wireless
networks: Issues and approaches. IEEE Personal Communications,

6(1):44--53, February 1999.
[4] J. S. Blogh and L. Hanzo. Third Generation Systems and Intelligent
Wireless Networking: Smart Antenna and Adaptive Modulation. John
Wiley & Sons, Ltd., 2002.
[5] S. Borst and P. Whiting. Dynamic rate control algorithms for HDR
throughput optimization. In Proceedings of IEEE INFOCOM'01, Anchorage,
Alaska, April 2001.
[6] Z. Fu, P. Zerfos, H. Luo, S. Lu, L. Zhang, and M. Gerla. The impact of
multihop wireless channel on TCP throughput and loss. In Proceedings
of IEEE INFOCOM'03, San Francisco, CA, March 2003.
[7] M. Grossglauser and D. N. C. Tse. Mobility increases the capacity
of ad hoc wireless networks. IEEE/ACM Transactions on Networking,

10(4):477--486, August 2002.
[8] P. Gupta, R. Gray, and P. R. Kumar. An experimental scaling law for
ad hoc networks. Technical report, University of Illinois at UrbanaChampaign,
 http://decision.csl.uiuc.edu/#prkumar, May 2001.
[9] P. Gupta and P. R. Kumar. The capacity of wireless networks. IEEE
Transactions on Information Theory, 46(2):388--404, March 2000.
[10] G. Holland and N. Vaidya. Analysis of TCP performance over mobile
ad hoc networks. In Proceedings of ACM MobiCom'99, Seattle, WA,
August 1999.
[11] G. Holland, N. Vaidya, and P. Bahl. A rate-adaptive MAC protocol for
multi-hop wireless networks. In Proceedings of ACM MOBICOM '01,

Rome, Italy, August 2001.
[12] F. Kelly, A. Maulloo, and D. Tan. Rate control in communication
networks: shadow prices, proportional fairness and stability. Journal
of the Operational Research Society, 49:237--252, 1998.
[13] R. Knopp. Achieving multiuser diversity under hard fairness constraints.
In Proceedings of IEEE International Symposium on Information Theory
(ISIT'02), Lausanne, Switzerland, June 2002.
[14] H. T. Kung and R. Morris. Credit based flow control for ATM networks.

IEEE Network, 9(2):40--48, March 1995.
[15] C. Li and E. Knightly. Coordinated multihop scheduling: A framework
for end-to-end services. IEEE/ACM Transactions on Networking,

10(6):776--789, December 2002.
[16] J. Li, C. Blake, D. DeCouto, H. I. Lee, and R. Morris. Capacity of ad
hoc wireless networks. In Proceedings of ACM MobiCom'01, Rome,
June 2001.
[17] X. Liu, E. Chong, and N. Shroff. Transmission scheduling for efficient
wireless network utilization. In Proceedings of IEEE INFOCOM'01,

Anchorage, Alaska, April 2001.
[18] H. Luo, S. Lu, and V. Bharghavan. A new model for packet scheduling
in multihop wireless networks. In Proceedings of ACM MobiCom'00,

Boston, MA, August 2000.
[19] L. Massoulie and J. Roberts. Bandwidth sharing: objectives and
algorithms. In Proc. of IEEE INFOCOM'99, New York, March 1999.
[20] J. Mo and J. Walrand. Fair end-to-end window-based congestion control.

IEEE/ACM Transactions on Networking, 8(5):556--567, October 2000.
[21] K. K. Mukkavilli, A. Sabharwal, E. Erkip, and B. Aazhang. On
beamforming with finite rate feedback in multiple antenna systems.

Submitted to IEEE Transactions on Information Theory, 2002.
[22] T. Nandagopal, T. Kim, X. Gao, and V. Bharghavan. Achieving MAC
layer fairness in wireless packet networks. In Proceedings of ACM
MobiCom'00, Boston, MA, August 2000.
[23] C. Perkins, editor. IP mobility support. RFC 2002, October 1996.
[24] D. Rajan, A. Sabharwal, and B. Aazhang. Delay bounded scheduling of
bursty traffic over wireless channels. IEEE Transactions on Information
Theory, 2003.
[25] A. Sabharwal. On the capacity of relay-assisted communications.
In Proceedings of IEEE GLOBECOM '02, Taipei, Taiwan R.O.C.,
November 2002.
[26] V. Tarokh, N. Seshadri, and A. Calderbank. Space-time codes for
high data rate wireless communication:performance criterion and code
construction. IEEE Transactions on Information Theory, 44(2):744--765,
March 1998.
[27] N. Vaidya, P. Bahl, and S. Gupta. Distributed fair scheduling in a
wireless LAN. In Proceedings of ACM MobiCom'00, Boston, MA,
August 2000.

