See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/221164105

Learning to share: Narrowband-friendly wideband networks

Conference Paper · January 2008

DOI: 10.1145/1402958.1402976 · Source: DBLP

CITATIONS READS
125 371

5 authors, including:

Nate Kushman Farinaz Edalat

Massachusetts Institute of Technology RKF Engineering Solutions, LLC
28 PUBLICATIONS   1,408 CITATIONS    6 PUBLICATIONS   313 CITATIONS   

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Farinaz Edalat on 01 June 2014.

The user has requested enhancement of the downloaded file.

 Learning to Share: Narrowband-Friendly Wideband
Networks
Hariharan Rahul, Nate Kushman, Dina Katabi, Charles Sodini, and Farinaz Edalat
Massachusetts Institute of Technology

ABSTRACT mance of narrowband devices in these bands, nor sacrifice their own
throughput or operating range. Overcoming this problem requires a
Wideband technologies in the unlicensed spectrum can satisfy the network design that achieves high throughput even when interferers
ever-increasing demands for wireless bandwidth created by emerging continuously exist, a fundamental departure from traditional wireless
rich media applications. The key challenge for such systems, however, networks, which are crippled by interference.
is to allow narrowband technologies that share these bands (say, This paper presents SWIFT, a Split Wideband Interferer Friendly
802.11 a/b/g/n, Zigbee) to achieve their normal performance, without Technology that safely coexists with narrowband devices operating
compromising the throughput or range of the wideband network. in the same frequencies. SWIFT’s key feature is cognitive aggrega-
This paper presents SWIFT, the first system where high-throughput tion: the ability to create high-throughput wireless links by weaving
wideband nodes are shown in a working deployment to coexist with together non-contiguous unused frequency bands that change as nar-
unknown narrowband devices, while forming a network of their own. rowband devices enter or leave the environment. Our design is moti-
Prior work avoids narrowband devices by operating below the noise vated by measurement studies [19, 27] showing that, while various
level and limiting itself to a single contiguous unused band. While wireless technologies exist throughout the spectrum, only a few such
this achieves coexistence, it sacrifices the throughput and operating technologies are usually operational in a house or small geographic
distance of the wideband device. In contrast, SWIFT creates high- area,1 and hence a large number of non-contiguous frequency bands
throughput wireless links by weaving together non-contiguous unused are likely to be unused. SWIFT’s ability to detect and utilize exactly
frequency bands that change as narrowband devices enter or leave these unoccupied bands, and compose them to build a single wireless
the environment. This design principle of cognitive aggregation link, allows wideband networks to operate at normal power without
allows SWIFT to achieve coexistence, while operating at normal affecting narrowband, and delivers on the promise of simultaneously
power, and thereby obtaining higher throughput and greater operating achieving high throughput, operating range, and coexistence.
range. We implement SWIFT on a wideband hardware platform, and SWIFT bridges two areas in wireless communications: cognitive
evaluate it in the presence of 802.11 devices. In comparison to a radios, and wideband and ultra-wideband design. While there has
baseline that coexists with narrowband devices by operating below been a lot of interest in cognitive communication, most proposals
their noise level, SWIFT is equally narrowband-friendly but achieves have focused on the licensed spectrum [12, 10, 16], where the pri-
3.6 − 10.5× higher throughput and 6× greater range. mary users of the band are known a priori, and hence this knowledge
may be incorporated into detecting if the band is occupied by the
Categories and Subject Descriptors C.2.2 [Computer Sys- known signal pattern. In contrast, SWIFT focuses on the unlicensed
tems Organization]: Computer-Communications Networks band, where narrowband devices are many, and their signal patterns
General Terms Algorithms, Design, Performance are unlikely to be known. Further, cognitive proposals attempt to find
a single unused band which they may opportunistically use, while
SWIFT aggregates the bandwidth of many such bands to maximize
throughput. Similarly to cognitive radios, Wideband (WB) and Ultra-
1 Introduction wideband (UWB) technologies have to cooperate with existing users
of the spectrum. They have, however, tried to bypass the coexistence
Users’ desires to share high definition audio and video around problem by reducing their transmission power below the noise floor
the home are driving the need for ever-increasing wireless band- of narrowband devices [34, 29, 4], and limiting themselves to a single
width [1, 9], and wideband radios, whose frequency bandwidth contiguous band. While this allows narrowband devices to oper-
spans hundreds of MHz to many GHz, have been proposed as a ate unhindered, it sacrifices the WB device’s throughput, operating
solution [9, 34, 20]. These radios mainly operate in the unlicensed distance, or both.
spectrum, which is populated by a variety of legacy narrowband To achieve its goal of high throughput, range, and narrowband-
devices (e.g., 802.11a/b/g, Zigbee), as well as a slew of emerging friendliness, SWIFT has to address three key challenges:
technologies (e.g., 802.11n). The key problem in operating these
wideband systems is to ensure that they neither hinder the perfor- • How does SWIFT detect the frequency bands that it must avoid, to
allow narrowband devices to operate normally? In the absence of
any information about the narrowband signal, traditional solutions
avoid frequency bands that show high narrowband power [10].
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
This approach uses observed power (or the lack of it) in a band
not made or distributed for profit or commercial advantage and that copies as a proxy for whether interference in this band is detrimental (or
bear this notice and the full citation on the first page. To copy otherwise, to irrelevant) to operation of the narrowband device, and is known to
republish, to post on servers or to redistribute to lists, requires prior specific have both false positives and false negatives [32]. Instead, SWIFT
permission and/or a fee. has a novel adaptive sensing technique that exploits common net-
SIGCOMM’08, August 17–22, 2008, Seattle, Washington, USA.
Copyright 2008 ACM 978-1-60558-175-0/08/08 . . . $5.00. 1
The measured average spectrum occupancy is 5.2% [19].
 work semantics, by observing that many unlicensed devices react power consumption, low-rate radios for precision location and track-
when faced with interference, either at the lower layers [7, 21], or ing systems, and high throughput radios for personal area networks
at higher layers [24]. This observation allows SWIFT to directly and wire replacement in homes and offices [9, 1].
address the key goal of cognition: identifying frequency bands An intrinsic problem for high-throughput wideband radios, how-
whose use could interfere with narrowband devices. Thus, SWIFT ever, is coexistence with narrowband devices with which they share
probes ambiguous frequencies, monitors the change in narrowband the unlicensed bands. Prior work tries to avoid interfering with nar-
power profile, and backs away if it perceives narrowband reaction. rowband devices by transmitting below their noise level [34, 29].
• How does the PHY layer operate across chunks of non-contiguous This approach inherently limits the throughput and operating range
frequencies? The current PHY layer of high-throughput wire- of the WB radio [34]. Further, in many cases, it fails to achieve its
less systems assumes a known and contiguous communication goal of protecting narrowband devices [29, 4]. Mishra et al. [28]
band, and breaks down in the presence of narrowband devices. propose to detect and avoid WiMax operating in the same band as
For example, even basic primitives like packet detection can be an ultra-wideband device. Their work however is specific to WiMax,
triggered incorrectly by power from narrowband transmissions. and can deal neither with general narrowband devices nor with a
SWIFT introduces a cognitive PHY that incorporates cross-layer dynamic environment. Also, their implementation considers only a
information from the adaptive sensing subsystem into the basic wideband sender and does not include a wideband receiver.
signal processing algorithms. While most prior work is focused on a single link and the PHY
• Given that different nodes might perceive different usable fre- layer, SWIFT’s components span multiple areas, including signal
quencies, how do SWIFT nodes communicate? Varying prox- processing, coding, and network protocols, which together success-
imity to narrowband devices between SWIFT transmitter-receiver fully address the issue of coexistence with dynamic and unknown
pairs may lead to differences in their choice of usable frequency narrowband devices.
bands. Since state of the art high-throughput wireless systems (e.g. (b) Cognitive Radios. The realization of the congested spectrum
OFDM) communicate across a frequency band by striping the data allocation and its inefficient utilization [19, 27] has led to a surge of
bits sequentially across sub-frequencies in the band, disagreement interest in cognitive communications. Work here has largely focused
in the set of usable sub-frequencies between a sender-receiver pair on detecting unused bands (spectrum sensing) and providing methods
leads to unknown insertions and deletions in the data stream, which for sharing these bands among cognitive radios (spectrum sharing).
cannot be dealt with by typical error-correcting codes. SWIFT’s Prior work on spectrum sensing focuses on the licensed band,
in-band consensus scheme transforms these insertions and dele- where it is crucial that cognitive secondary users do not interfere
tions into bit errors, which can be dealt with using standard error- with the licensed primary user. The most basic approach involves
correcting techniques, and hence enables communication despite measuring the energy level in a band. Energy detection is cheap, fast,
uncertainty in the environment. and requires no knowledge of the characteristics of the signal. How-
We have built SWIFT in a custom wideband radio hardware [20]. ever, choosing energy thresholds is not robust across a wide range of
Our implementation addresses the major details of computational SNRs [10]. Though more sophisticated mechanisms such as matched
complexity, storage, and pipelining inherent in building a wideband filter detection [10] are more accurate, they require knowledge of
wireless transceiver and apparent only at the hardware level. We the transmitted signal (modulation, packet format, pilots, bandwidth,
evaluate our design in a testbed of wideband nodes and 802.11 nar- etc.) and thus work only for known technologies.
rowband devices. Our results reveal the following findings. Architectures for spectrum sharing fall in two categories: central-
ized and distributed [10]. Centralized approaches [3, 15, 14] require
• SWIFT safely coexists with narrowband devices while simultane- a controller, such as a base station or spectrum broker, to allocate
ously providing high throughput and good range. In comparison to spectrum to all cognitive users. Distributed approaches [35, 36, 12,
a baseline system that coexists with narrowband devices by operat- 16, 25] have MAC protocols that rely on one or more control channels
ing below their noise level, SWIFT is as narrowband-friendly, but to coordinate spectrum access.
its throughput is 3.6 − 10.5× higher, and its range is 6× greater. While our work builds on these prior foundations, it makes three
• Adaptive sensing is effective. As compared to a threshold based major departures. First, cognitive radios focus on finding a single
approach, which is neither efficient for wideband nor safe for nar- contiguous unoccupied band, whereas SWIFT weaves together mul-
rowband across all locations, adaptive sensing accurately identifies tiple non-contiguous unoccupied bands to create a high-throughput
interfered frequency bands, and provides efficiency while still wideband link. Second, SWIFT introduces new spectrum sensing
being safe for narrowband. mechanisms that exploit network semantics to strengthen traditional
• SWIFT nodes can communicate despite disagreement over narrow- energy based techniques for unknown signals. Third, SWIFT allows
band spectrum usage and tolerate up to 40% disagreement about communicating nodes to agree on usable frequencies using a fully
the usable frequency bands. distributed consensus scheme that requires no control channels.
To the best of our knowledge, SWIFT is the first system where
wideband nodes are shown in a working deployment to coexist safely 3 Problem Domain
with unknown narrowband devices, while forming a network of their
own. SWIFT is designed to provide high throughput wireless connectiv-
ity for rich media appliances in a home scenario. It operates in the
2 Related Work unlicensed spectrum, and is intended to function in the presence of
narrowband devices that utilize the same part of the spectrum, and
SWIFT brings together research in two threads of wireless communi- which might persist for long periods, or arrive and depart within min-
cations: wideband systems, and cognitive radios. utes or hours, e.g., a laptop utilizing an 802.11 wireless connection.
(a) Wideband Systems. The last couple of years have seen SWIFT is a cognitive architecture for OFDM wideband radios.
tremendous successes in the implementation of WB and UWB ra- We focus on Orthogonal Frequency Division Multiplexing (OFDM)
dios [18, 20, 9, 34]. This work falls in two major categories: low because it has emerged as the technique of choice for the majority of
 OFDM Symbol
In Frequency Domain
BIN 1 BIN 1
BIN 2 BIN 2

Demodulation
Modulation
WB Radio Adapter
… 011010 …

IFFT
… 011010 …

FFT
DAC ADC
WB Radio Adapter
Appliance
BIN W BIN W

OFDM Transmitter Side OFDM Receiver Side Appliance

WB Radio Adapter WB radio Adapter

Figure 1: Schematic of an OFDM System

Very High Rate
Appliance
Appliance
wireless technologies, such as wideband digital communication [20],
ultra-wideband [5], 802.11 a/g/n [7, 8] and WiMAX [11]. The rest of
our description focuses on single antenna radios, but our ideas are also Bands occupied
Non-contiguous
by narrowband
applicable to wideband MIMO radios, as they too use OFDM [13]. devices bands usable by
wideband
Robust detection of narrowband devices without any knowledge of
their signal patterns or other characteristics is impossible [32]. Since
it is impractical to assume known signal patterns in the unlicensed
band, SWIFT focuses its design on the practical scenarios that could Spectrum
arise in the environment of interest. Specifically, SWIFT addresses Figure 2: Cognitive Aggregation: While narrowband devices exist
situations in which the following constraints apply: (e.g., 802.11 laptop), SWIFT still uses the remaining non-contiguous
1. It is acceptable to treat narrowband traffic as best effort. Specifi- chunks of spectrum as if they were one wireless link.
cally, narrowband devices should continue to experience the same
average throughput and loss rate in the presence of wideband nodes 5 SWIFT
as without them, but their requirements are not any more stringent
than what is expected from today’s wireless LANs. SWIFT is designed around the concept of cognitive aggregation. Sim-
ilar to the cognitive radio vision, cognitive aggregation is based on
2. The capacity of the wideband network exceeds its peak traffic. detecting narrowband systems and avoiding their frequency bands.
This implies that the medium exhibits frequent idle intervals such Unlike prior cognitive systems, which use only a single contiguous
that narrowband devices that perform carrier sense are not com- band, cognitive aggregation merges many non-contiguous bands into
pletely locked out. Sufficient capacity can arguably be obtained by a single high-throughput communication channel, as shown in Fig. 2.
increasing the spectrum width spanned by the wideband radio. Such a design is critical when using a wide band in the unlicensed
3. Narrowband technologies of interest in this paper react to interfer- spectrum since a wide contiguous unused band typically does not
ence. This reaction can be at lower layers, for example, carrier- exist. SWIFT implements a cognitive aggregation design by utiliz-
sense abstaining from using the medium, or autorate changing ing three key components: (a) a spectrum sensing mechanism based
modulation schemes, or at the higher layer, for example, TCP on determining how SWIFT’s selection of frequency bands impacts
backing off on sustained packet loss. Further, these devices are narrowband transmissions, rather than just how the narrowband trans-
expected to operate at reasonable SNRs (a few dB above the noise missions look to SWIFT, (b) a cognitive PHY layer that can operate
floor, e.g. 802.11a/b/g/n). Narrowband devices that operate below over non-contiguous spectrum bands, and (c) a consensus protocol
or around the noise floor are expected to have their own mecha- that allows SWIFT nodes to agree on usable frequency bands despite
nisms to combat interference, as they need them in such a regime. uncertainty about which bands are occupied by narrowband devices.
Below, we explain each of these components in detail.

4 OFDM Background 5.1 Adaptive Spectrum Sensing

This section provides a simplified description of OFDM focused only SWIFT senders must learn the set of OFDM bins in which they can
on issues related to this paper. OFDM divides the used RF bandwidth send while being narrowband-friendly.
into many narrow sub-channels, called OFDM bins. Each OFDM bin
can be treated independently from other bins, and may use a different 5.1.1 How do we detect bins that interfere with narrowband?
modulation (e.g., BPSK, 4-QAM) or transmission power. A data
stream is striped into bits, with different numbers of bits assigned Ideally, SWIFT could directly measure how its choice of transmit
to each bin based on its modulation scheme. An assignment of bins affects a narrowband device. Since this is typically not possible,
modulated bits to each of the OFDM bins is called an OFDM symbol, and given that one does not know the signal details for arbitrary unli-
see Fig. 1. The frequency representation of the OFDM symbol is censed narrowband devices, prior cognitive devices passively listen
converted to a time domain OFDM symbol by using an Inverse Fast for narrowband devices, and avoid all frequency bins in which they
Fourier Transform (IFFT) and sent on the medium by the transmitter. see power above some threshold [10]. This approach essentially uses
The receiver first determines the exact sample at which the packet information about how SWIFT observes the narrowband transmis-
starts. It then aligns the time samples on OFDM symbol bound- sions to guess how a SWIFT transmission would be observed by the
aries, and performs a few basic signal processing tasks like Carrier narrowband device. Such an approach is problematic for two reasons.
Frequency Offset (CFO) and channel estimation. Next, the aligned First, it is difficult to pick a power threshold [32] to precisely iden-
time signal is passed to a Fast Fourier Transform (FFT) module to tify occupied bins, because the correct value varies with time and
produce the frequency representation. The data symbols are then proximity to the narrowband device. Fig. 3 illustrates this issue. It
converted to their frequency representation, corrected for the channel, shows the power profile of an 802.11a narrowband device operating
and demodulated to retrieve the transmitted data bits. on channel 52, as observed by two SWIFT nodes at different dis-
 200 5.1.2 Detecting Narrowband Reaction
Observed 802.11a Power Distant SWIFT
180 Best Distant Threshold
160 Nearby SWIFT
SWIFT continuously senses the medium whenever it is not sending
140 Best Nearby Threshold or receiving a packet. It converts the incoming time signal to the
120 frequency domain using an FFT, and then calculates the current
100 power in each bin. These power measurements are used both to
80
60 detect the existence of a narrowband device, and to identify whether
40 the narrowband device has reacted to the wideband device.
20 SWIFT detects the presence of a narrowband device in a bin, by
0 comparing the power in that bin to the noise floor. SWIFT computes
-60 -40 -20 0 20 40 60
the noise floor by taking advantage of its wide band. Since it is highly
Figure 3: 802.11a Power Profile: The observed power of an unlikely that narrowband devices are simultaneously present in all
802.11a transmitter at different SWIFT locations is very different, bins, SWIFT just computes the minimum power across all bins and
highlighting the difficulty in picking a power threshold that works at averages it over time to estimate the noise floor. Before SWIFT runs
all locations. its adaptive sensing algorithm to choose the correct set of bins, it uses
a conservative threshold that declares a bin narrowband-occupied if
the power in that bin exceeds the noise floor by 3 dB in any sample,
tances from the 802.11a transmitter. In this scenario, the narrowband and narrowband-free otherwise. A sample is considered narrowband-
device uses bins 3 through 23. Clearly, no single fixed threshold occupied if any bin in that sample is narrowband-occupied.
would eliminate exactly the correct set of bins used by the narrow- SWIFT also uses its power measurements to compute four metrics
band device at both locations. This problem becomes even worse that capture the most common responses to interference.
in the presence of variable power levels among narrowband devices.
For example, portable 802.11 devices such as laptops and handheld • Inter-transmission time captures the behavior of narrowband de-
devices often transmit at power levels well below the maximum in vices that react to interference by backing off (e.g. 802.11 or TCP
order to conserve their battery, meaning that even though SWIFT’s backoff). It is computed by counting the number of consecutive
effect on two devices in different locations might be very different, narrowband-free samples.
the transmissions from those two devices might be indistinguish- • Transmission duration captures the behavior of devices that fall
able from SWIFT’s perspective. Accounting for all this variability back to more robust, lower rate modulation schemes, thereby tak-
requires using a very conservative threshold that wastes many bins. ing a longer amount of time for each transmission (e.g. autorate
in 802.11). It is computed by counting the number of consecutive
Second, even if one could identify the exact bins the narrowband narrowband-occupied samples.
device uses for its transmissions, this may not be the correct set of • Average narrowband power allows SWIFT to deal with multiple
bins to avoid. Since transmitters leak power into bins adjacent to narrowband devices in the same band (e.g., two 802.11 devices).
the ones they use, a wideband transmitter might need to avoid bins If SWIFT interferes with a nearby device causing it to backoff, but
that are unused by the narrowband device if using them would leak a more distant device fills in the freed bandwidth such that none
significant power into the narrowband bins. Conversely, a wideband of the other metrics changes, the average power will significantly
device might be able to use bins that are used by the narrowband decrease, allowing SWIFT to detect the change. This metric is
device without affecting narrowband operation. This might happen if computed by averaging the power in narrowband-occupied samples
the narrowband device is far away from the wideband transmitter, or over a window.
uses highly redundant coding schemes (e.g., Zigbee [37]). Because • Probability of transmission immediately after SWIFT captures
these effects depend on the distance and receive sensitivity of the whether SWIFT triggers the carrier-sense reaction of narrowband.
narrowband device, it is impossible to account for them without being If SWIFT triggers narrowband carrier-sense, the narrowband de-
extremely conservative in the choice of threshold. vice will not transmit immediately after a SWIFT packet, because
The key problem with current solutions is that they use the wide- it waits to ensure that the medium is free (In 802.11, this translates
band device’s view of the narrowband transmissions in an open loop, to the DIFS, followed by a random contention window). The met-
as a proxy for how the narrowband device will observe the wideband ric is computed by looking at the power immediately after SWIFT
transmissions. Asymmetric links, and varying transmission powers finishes transmitting a packet, and setting a flag to 0 if the sample
and receive sensitivities, make this a poor proxy. SWIFT instead is narrowband-free, and 1 otherwise. The probability is computed
uses a technique we call adaptive sensing, which closes the loop by as the average of these flags over a recent window.
taking advantage of the observation that many narrowband devices SWIFT maintains sufficient statistics to compute the mean and
react in some perceivable way if wideband transmissions disrupt variance of each metric. To achieve high confidence in the value of
their transmissions. In particular, a large class of narrowband tech- a particular metric, SWIFT needs to collect multiple measurements
nologies in the unlicensed spectrum reacts to interference, either at of that metric. Note that for the first three metrics, SWIFT gets one
lower layers (e.g., carrier-sense and autorate) or higher layers (e.g., measurement every time it sees a narrowband transmission. The last
TCP or end-user backoff). Intuitively, SWIFT pokes the narrowband metric is different, however, in that it can be measured independent
device by putting power in ambiguous bins, notes any changes in of whether the narrowband device transmits or not. If the narrow-
the narrowband power profile, and backs away if such a reaction is band device has nothing to send though, the fact that no narrowband
observed. transmission is observed provides no information. Hence, SWIFT
Note that our goal with adaptive sensing is not to use narrowband only includes samples of this metric when it senses a narrowband
bins during short gaps in narrowband transmissions; rather, we design transmission within some maximum time after a SWIFT packet (1
it to immediately relinquish bins that it suspects of being used by ms in our implementation). Thus, the confidence of our estimates of
narrowband devices, and reuse them only when confident that the all four metrics depends only on how many samples are obtained, and
narrowband devices have disappeared for several minutes. is independent of how sporadically the narrowband device transmits.
 BIN 1
Narrowband Compute Advance Compute BIN 2
Detected yes Back off to Are
no
START Conservative Normal UsableBins Current Metrics

Modulation

Inverse FFT
Above Noise Threshold Metrics by 1 Metrics Normal? DAC
Floor? … 01100010 …
no TIME SAMPLES OFDM SYMBOL
yes
All Bins
no Marked
Usable? BIN W
yes
yes
Continue to Retreat to
no Are Metrics

FREQUENCY
Compute Previous BIN 1 BIN 1 BIN 1

DOMAIN
Normal? Choice
Metrics BIN 2 BIN 2 BIN 2

DO
FRE
BIN W BIN W BIN W
Figure 4: Control Flow for Adaptive Sensing Algorithm UNALIGNED FFT ALIGNED FFT

5.1.3 Adaptive Sensing Algorithm TIME DOMAIN

We define a bitvector UsableBins, which identifies the set of bins Figure 5: Conversion of bits into OFDM symbols: Values in indi-
that SWIFT currently uses. The adaptive sensing algorithm starts vidual frequency bins are combined in each time sample, and can be
with a conservative choice of UsableBins that does not interfere recovered only by computing appropriately aligned FFTs.
with the narrowband device, and iteratively tightens the setting of
UsableBins to converge on the maximal set of usable bins that opportunity to transmit during SWIFT’s idle intervals, and hence be
does not affect the narrowband device. Fig. 4 shows the control flow quickly detected, allowing SWIFT to immediately back away and
of our algorithm, which we describe in detail below. trigger the adaptive sensing algorithm for this new narrowband group.
Whenever SWIFT first detects narrowband power in a bin (using
the conservative threshold), it immediately backs away from that bin, 5.1.4 Measuring Statistically Significant Changes
and updates UsableBins accordingly. This conservative choice
When should SWIFT decide that changes in some metric are not due
of UsableBins allows SWIFT to be confident that observations
to statistical aberrations, but reflect a real change in the performance
made in this state represent normal narrowband behavior.
of the narrowband device?
After gathering enough data at this normal setting, SWIFT begins
SWIFT uses a statistical test called a t-test, typically used to decide
the process of determining a choice of UsableBins that does not
whether a drug has had a statistically significant effect on the popula-
affect the narrowband device, but provides a maximal number of
tion studied [17]. A t-test takes the means, variances, and number of
available bins. It starts by grouping contiguous sets of narrowband-
samples of the two compared sets: normal and current. It computes
occupied bins into a single narrowband group. Each narrowband
the following t-value where the x̄ ’s and σ ’s represent the means and
group is then assigned a top and bottom bin which bound, for this
standard deviations, respectively, of the two sets, and nnorm and
narrowband group, the range of bins which must be left unused.
ncurr refer to the number of samples in the normal and current set:
Next, SWIFT will try to grow UsableBins by using the top and
bottom bins in each narrowband group and observing whether the x̄norm − x̄cur
t=q .
narrowband device reacts. At each step, SWIFT alternates between σnorm σcur
nnorm + ncur
reducing the top bin by one and increasing the bottom bin by one. For
each choice of UsableBins, SWIFT waits to gather data measuring To determine whether any difference between the means is statisti-
the effect of this new choice. It continuously monitors the incoming cally significant, the t-value must be combined with an alpha level,
data by comparing the metrics with this bin choice to those observed which represents the acceptable probability of being wrong. In our
under normal behavior with the conservative bin choice. If, at any case, this value represents the probability that the t-test will tell us
point, SWIFT determines that it has impacted any of the metrics, it that SWIFT is interfering even if it is not. This is a parameter which
immediately moves back one step, and resets UsableBins to the effectively sets the aggressiveness of SWIFT. We use an alpha level of
previous decision. If, however, after gathering enough data, SWIFT 0.05, typical for scientific and medical studies. The t-value combined
determines that none of the metrics are impacted, it moves on to the with the alpha level and the total number of samples is then used in
next step, and tightens its choice further by one bin. a table look-up to determine whether the t-test passes, i.e., whether
For each narrowband group, SWIFT independently continues this SWIFT has had a statistically significant impact on narrowband.
process until it either reaches a bin choice for which it notices the
narrowband device reacting, in which case it retreats to the previous
5.2 Cognitive PHY
UsableBins setting, or it marks as usable all bins in this nar-
rowband group and still notices no reaction. At this point, SWIFT The cognitive PHY uses the output of adaptive sensing to provide a
continues to monitor the metrics and compare them to normal. If single high-throughput link over the set of usable bins.
it notices a change at any point, SWIFT retreats to the conservative On the transmitter, this means ensuring that no power is used in
choice of UsableBins, recomputes normal metrics, and repeats bins marked as narrowband-occupied by the adaptive sensing module.
the probing process, as shown in Fig. 4. This is straightforward with OFDM since it naturally allows different
Note that this algorithm inherently deals with dynamics. For ex- power assignments for each frequency bin.
ample, if the narrowband device moves closer or farther after SWIFT On the receiver side, the cognitive PHY has to ensure that the
has finalized a bin choice, the average narrowband power metric will receiver can receive in non-contiguous bins even when narrowband
change from normal, and cause SWIFT to reinitiate the entire probing devices are using the other bins. At first, it might seem that this can be
process. Furthermore, if all narrowband devices in a group depart, done analogous to the transmitter by taking the FFT of the incoming
SWIFT will stop seeing any transmissions in the narrowband group, signal, and just using values from the bins of interest. However,
time out the entire group after a predefined interval, and reclaim this is impractical. To understand why, consider the frequency-time
these bins. Also, as articulated in §3, a narrowband device appearing diagram in Fig. 5 which illustrates how the N OFDM frequency bins
in a new band currently occupied by SWIFT will always have the are converted to N time samples that together represent an OFDM
symbol. As can be seen, the correct frequency domain values can
be retrieved from the time samples only when the FFTs are aligned Bin Index 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5
correctly on OFDM symbol boundaries. But the receiver can align Symbol 1 Symbol 2 Symbol 3
the FFT correctly on symbol boundaries only if it knows the starting
sample of a packet in the first place! Transmitter 0 1 1 0 1 1 0 1 0 1 0 0
Hence, we need to modify a few basic receiver algorithms to cope
with non-contiguous bands.
Receiver 0 X 1 1 0 1 Y 1 0 1 0 Z 1 0 0
(a) Receiver Packet Detection: In order to perform any processing
on a packet, the receiver first needs to determine the start of the packet
Figure 6: Bin Disagreement Causes Communication Failure: If
within a few time samples. Typically, this is done using the double
the transmitter sends in bins 1, 3, 4, and 5 while the receiver listens
sliding window approach [23], which uses energy ratios to determine
in 1, 2, and 5, then the receiver will decode noise in bin 2 as data, and
the time sample where a burst of energy is received on the medium.
miss data in bins 3 and 4. These insertions and deletions will cause
Since this operation happens in the time domain, it cannot dis- a misalignment in the demodulated data stream, creating an error
tinguish between energy from narrowband devices and wideband pattern than cannot be rectified by standard error-correcting codes.
transmitters, and can be spuriously triggered by narrowband transmis-
sions. Recall that SWIFT concurrently transmits with narrowband
devices by using separate frequencies. Hence, if the receiver is kept proximity to narrowband devices and variations in time make it likely
busy with false packet detections, it is very likely to miss desired that a transmitter and receiver identify different bins as usable. For
wideband transmissions.2 example, a wideband sender and receiver that are just a few meters
The solution is to actively filter the narrowband devices, allowing apart may differ in their perspectives of narrowband-occupied bins
the receiver to perform packet detection on the clean signal consisting by as much as 10-20 MHz as we show in §7.2.
primarily of power from wideband transmitters. The choice of the This disagreement between a transmitter and its receiver can be
bins to filter is driven by the adaptive sensing module. However, a fatal obstacle to establishing an OFDM communication link. To
the receiver may not be able to use a filter per narrowband group understand why, recall that an OFDM transmitter stripes data across
since filters are resource-intensive in hardware. Hence, SWIFT is all usable OFDM bins. A receiver reconstructs the original data by
designed to use a small fixed number of bandstop filters, whose extracting bits from the individual bins. Thus, as shown in Fig. 6, if
widths and center frequencies are dynamically configured. Note that the receiver expects data in a bin that the transmitter did not send in, it
since these filters are purely on the receiver side, by definition, they will result in insertion of bits into the data stream. Conversely, if the
do not affect narrowband devices. A particular filter choice that is transmitter sends data in a bin that the receiver does not expect data in,
not perfectly aligned with the desired set of bins to be filtered only it will manifest itself as deletions of bits from the data stream. Thus,
affects packet detection to the extent of the amount of narrowband disagreements about bins result in alignment and framing errors, and
energy that it lets in, or the amount of wideband transmitted energy it produce a wireless channel that has unknown insertions and deletions,
filters out. The filter computation problem is formulated as a dynamic which conventional error correcting codes cannot deal with.
program that eliminates as many narrowband bins as possible, while We solve this problem using two mechanisms: (a) an infrequent
maximizing the amount of received wideband energy. The details of synchronization phase when the communicating wideband pair has
this optimization are omitted here for space, but described in [31]. a drastic disagreement, say, when a wideband node boots up, or
when many narrowband devices in different bands appear simultane-
(b) Receiver Packet Processing: Now that the start of the packet ously, and (b) a low overhead handshake, which is used when nodes
has been detected accurately, the receiver has the right alignment that have previously agreed experience a limited disagreement, say,
for the symbols and the rest of the packet processing can be done because a single narrowband device was turned on or moved closer.
in the frequency domain over the actual bins used by the wideband
SWIFT nodes are equipped with a robust initial synchronization
system. Specifically, carrier frequency offset estimation, which is
mechanism. Each SWIFT node divides the whole transmission
traditionally done in the time domain, is instead performed in the fre-
band into chunks of 16 bins, checksums and codes the value of
quency domain after zeroing out the contributions of bins occupied by
its UsableBins, and sends it simultaneously in all chunks. Assum-
narrowband, as determined by adaptive sensing. This permits a more
ing that the bandwidth of the wideband node is large enough, and has
precise estimate than an application of the time domain estimation
enough bins that are not interfered with narrowband, at least one of
algorithms on the noisy filtered signal used for packet detection.
these chunks in this sync packet will be received correctly, allowing
(c) Data reception: Recall that the transmitter, while assigning data the nodes to establish connectivity. Note that the sync packet uses all
to bins, zeros out all bins that are deemed unusable by adaptive OFDM bins, and hence does not suffer from an alignment problem.
sensing, and stripes data only across the remaining bins. Similarly, Even after a SWIFT node pair is synchronized, they can still suffer
when the receiver collects the received data, it only utilizes bits from from occasional disagreements, for example, when adaptive sensing
bins that are deemed unoccupied by narrowband devices. Again, we changes the set of usable bins on a node. We leverage the existing
note that since data reception happens after the alignment provided agreement to transform the potential disagreements into bit errors, i.e.,
by packet detection, it can work on the unfiltered signal and hence we transform the hard problem of unknown insertions and deletions
can precisely remove bins susceptible to narrowband interference. into the simpler problem of bit errors, a problem that all wireless links
know how to deal with by adding practical error correcting codes.
5.3 Communication Over Uncertain Bands To do so, SWIFT exploits the following key observation. If the
transmitter stripes the data across the previously agreed bins, there
Since each node in a SWIFT network independently decides the will be no deletions or insertions. The problem, however, is that,
bands that it can use for transmission and reception, differences in by transmitting in the old bins, some of which may no longer be
2
Due to the hardware pipelining typical to receivers [23], they cannot receive packets free, the transmitter might hinder a narrowband device. To address
while they are still working on the spuriously detected packet and have not rejected it. this problem, SWIFT stripes the data across the previously agreed
 Parameter Value
Carrier Freq. 5.247 GHz This is in contrast to current networks where a node decodes received
Data BW 100 MHz packet headers to determine if they are intended for itself.
Number of Bins 100 (×1 MHz)
Symbol Period 1.4 µ s SWIFT adapts the technique of correlation with known
Uncoded BER 10−3 pseudonoise sequences, typically used for packet detection, to de-
Bin Modulation BPSK, 4-
16-, 64-QAM velop a solution at the link layer. It is well known that pseudonoise
Max Link Len 10 m sequences exhibit low correlation with each other while showing high
Avg. Output Pwr 7.5 dBm
correlation with themselves, thereby allowing identification of spe-
cific pseudonoise sequences purely by correlation [30]. Transmitter
MAC addresses in SWIFT are pseudonoise sequences, and appear in
Figure 7: Wideband Radio Used in SWIFT a known and fixed symbol location in the received packet. When a
receiver detects a packet, it correlates it against its neighboring nodes’
MAC addresses to determine the transmitter, and hence the set of
bins, but transmits only in the subset that is still usable. The receiver, bins. This requires a receiver to maintain a table of neighbor MAC
which still expects to receive data across the old agreement, receives addresses; a receiver learns about a neighbor’s MAC address during
data in the intersection of the old and new bins correctly, but sees the initial sync packet where they exchange their mutually usable set
errors in the other bins. However, this can be easily fixed by using a of bins. Note that receiving the sync packet itself does not require
simple error correcting code with sufficient redundancy to cover the prior bin agreement, as described in §5.3.
expected extent of disagreement between old and new bins.
SWIFT uses a low-overhead handshake to quickly resolve dis-
agreements. The data in the handshake is the new set of usable bins, 6 Implementing SWIFT
and the striping technique is as described above. Once the handshake
We have implemented SWIFT in a custom wideband radio transceiver
terminates, the nodes resume normal data exchange.
platform developed by the WiGLAN research project [20]. The
WiGLAN transceiver board, shown in Fig. 7, connects to the PC via
5.4 Network Issues the PCI bus, and acts like a regular network card. The transceiver [26]
This section briefly describes how we compose multiple SWIFT links consists of three parts: 1) the RF front-end, which captures the analog
to build a network. signal, 2) the data converters, which convert between analog and dig-
ital, and 3) the digital baseband modem. All digital processing, such
(a) The MAC: We use a carrier sense based MAC similar to as packet acquisition, channel estimation etc., is done in baseband.
802.11 [22]. A node senses the medium and transmits if the medium Our prototype has two components: the driver and the firmware.
is not busy. However, a direct application of the carrier sense tech- The former is implemented in software, and the latter in FPGA.
nique of narrowband radios, which just checks for the total received Driver: The driver presents a standard network interface to the
power in the band to exceed a threshold, will unnecessarily reduce the kernel. In addition to this typical functionality, the driver offloads
transmission opportunities of SWIFT nodes since narrowband trans- from the FPGA any computation that is too complex for hardware
mitters are always likely to be using some part of the band and hence and is not on the critical path of an OFDM symbol. For example, the
preventing the wideband radio from transmitting. Instead, SWIFT’s driver implements the metric computation and t-test (§5.1). Our cur-
carrier sense focuses only on the bins declared usable by adaptive rent prototype implements two metrics: average narrowband power,
sensing. Specifically, when a node wants to send, it computes an FFT and probability of transmission immediately after SWIFT.
of the observed power, and proceeds with its transmission only if a Firmware: Several of SWIFT’s major components that need to be
large fraction of its usable bins are below the wideband carrier sense on the critical path, such as narrowband power measurement (§5.1),
threshold.3 Further, while wideband nodes can use an 802.11-like the cognitive PHY (§5.2), the band consensus protocol (§5.3), and
MAC, they need to wait for a relatively longer period to check that the MAC (§5.4), are implemented on the FPGA. We design SWIFT’s
the medium is idle, i.e., they should use a longer DIFS interval than algorithms in the Simulink environment, which has a hardware model
typical values picked by narrowband devices. This ensures that a for the Xilinx Virtex-4 SX35 FPGA that we use. The code is then
narrowband device that has just arrived into the environment can compiled into an intermediate form using Xilinx tools [6]. We use
quickly access the medium and trigger adaptive sensing. Verilog to integrate this intermediate form with the PCI subsystem,
The SWIFT MAC randomly jitters the start of a probing epoch to and create the final hardware representation of our code.
ensure that different SWIFT nodes perform adaptive sensing indepen-
dently. Further, a node uses control packets analogous to RTS/CTS
to notify other SWIFT nodes of the start and end of a probing epoch 7 Performance Evaluation
in order to avoid simultaneous probing by multiple nodes. While
We evaluate SWIFT in a 12 node testbed consisting of four wideband
this solution works for small wideband networks, extensions to larger
nodes, and eight 802.11a nodes. Fig. 8 shows the experimental envi-
networks may require more sophisticated mechanisms to leverage
ronment, which has high diversity due to the presence of walls, metal
probing results across multiple SWIFT nodes.
cabinets, desks, and various non-line-of-sight node locations. The
(b) Transmitter Identification: The alert reader might have ob- exact choice of node locations for each experiment will be described
served that a SWIFT receiver potentially needs to receive and decode along with the results for that experiment.
packets from multiple transmitters; however, decoding a packet re- Wideband Devices. We use the WiGLAN wideband hardware de-
quires knowledge of the exact set of mutually agreed bins over which scribed in §6, whose specifications are in Fig. 7. It has 100 OFDM
the data is striped, and this mutual agreement is likely to be different data bins, numbered from -50 to +50, with bin 0 never being used. For
with different transmitters. Hence, the SWIFT receiver needs to iden- all schemes, the wideband devices are evaluated while continuously
tify the transmitter of a packet even before it can decode the packet. sending 10 ms packets with a 1 ms gap between packets.
3
Note that the objective of wideband carrier sense is not to correctly decode the Narrowband Devices. These nodes run 802.11a in channel 52, cor-
received signal, but rather to measure received power, which does not require alignment. responding to wideband bins 3 through 23. 802.11a nodes send UDP
 40 600
802.11a Throughput (Mbps)

NORM Throughput (Mbps)

35 NORM (No 802.11a)
500
30
NORM (With 802.11a)
25 400
20 300
15
200
10
5 100
0 0
LOW NORM SWIFT No Wideband 1 2 3 4 5 6 7 8
Compared Scheme Location
(a) Throughput of 802.11a (b) Throughput of Non-Adaptive Wideband (NORM)
160 600

SWIFT Throughput (Mbps)

LOW Throughput (Mbps)

140 LOW (No 802.11a) SWIFT (No 802.11a)

500
120 LOW (With 802.11a) SWIFT (With 802.11a)
100 400
80 300
60
200
40
20 100
0 0
1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8
Location Location
(c) Throughput of Low-Power Wideband (LOW) (d) Throughput of SWIFT
Figure 9: Approaches to Narrowband-friendliness: Presents the throughput-range tradeoff, and shows that SWIFT, illustrated in (d), is as
friendly to 802.11a as LOW, while attaining dramatically higher throughput and operating range.

7.1 Throughput and Range

This experiment explores if it is possible to be as narrowband-friendly

as a transmitter operating below the noise level, while preserving the
good throughput and range of a normal-powered wideband system.
Method. We place the wideband transmitter in location tx, and test
its performance to the wideband receiver which is placed in each of
locations 1 through 10. For each location, we measure the throughput
of LOW, NORM, and SWIFT with and without interfering 802.11a
traffic, and plot the results in Fig. 9.
Results. Fig. 9 demonstrates that, while both NORM and LOW are
flawed, SWIFT can deliver on the fundamental goal of simultaneously
Figure 8: Testbed Map: Node Locations are Highlighted. achieving the high throughput and wide range of NORM, while being
as narrowband friendly as LOW. In particular, we see that:
streams at the highest rate supported by the medium, except for ex-
periment 7.5, in which they use TCP. The protocol, signal details,
and occupied bands of 802.11a are, of course, unknown to SWIFT. • Throughput and range of LOW are limited: Fig. 9(c) shows
Compared schemes. We compare the different schemes by config- that LOW fails to get any throughput after location 2, and has
uring our wideband hardware to run one of: 3.6 − 10.5× lower throughput than SWIFT and NORM.
• NORM is not narrowband friendly: We can see from Fig. 9(a)
• SWIFT: This is the SWIFT protocol implemented as in §6. that NORM significantly reduces 802.11a throughput.
• Low-power wideband (LOW): This is a baseline system that • SWIFT has high throughput and range: From Figs. 9(b) and
operates below the noise level to avoid interfering with narrowband 9(d), we can see that in all locations, SWIFT achieves the same or
devices. Specifically, it transmits signals with a power spectral greater throughput than NORM, with or without 802.11a.
density of -41 dBm/MHz, the FCC maximum for UWB devices [2]. • SWIFT is narrowband friendly: From Fig. 9(a), we can see that
• Non-adaptive wideband (NORM): This is a system that trans- 802.11a throughput is unaffected by SWIFT.
mits across a wide band at the normal power of our hardware
platform, but does not adapt to narrowband devices. We see from Figs. 9(b) and 9(d) that SWIFT surprisingly achieves
Note that both LOW and NORM will suffer drastic bit errors in bins higher throughput than NORM in the presence of 802.11a. This is
used by 802.11a when it is turned on. For conservative comparison because SWIFT intelligently avoids 802.11a occupied bins, while
in this case, we therefore consider idealized versions of these systems NORM uses these bins, suffers errors due to high narrowband power,
that use the minimal amount of coding required to correct these errors. and hence incurs additional overhead to correct errors in these bins.
 600 20
20

Number of Wasted Bins

Inefficient!

Number of Wasted Bins

500 15
15
Inefficient!
Ideal Threshold

400 10
10

300 5
5

200 0
0
30
Unsafe!

(Mbps)
30

(Mbps)
100 25

INF
INF
INF
INF
25

802.11aa Throughput
20

802.11aa Throughput
20
0
15 15
1 2 3 4 5 6 7 8 9 10

Th

Th
10 10

Location 5 5
0 0
Figure 10: No Single Threshold Works Across Locations: This 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10

figure plots the ideal threshold that ensures safe narrowband oper- Location Location

ation while maximizing bins usable by wideband. 802.11a nodes (a) Median Threshold (b) Minimum Threshold
at locations 7-10 are not affected by wideband, and hence the ideal
Figure 11: No Threshold is Safe and Efficient in All Locations
threshold for these locations is infinity.
25
7.2 Power Threshold Sensing 21 21 21 SWIFT

Number of Unusable Bins

20 Correct
In §5.1, we discussed the intractability of a threshold based algorithm. 17
Here, we present results validating that claim, first showing the dif- 15
15 13
ficulty of picking a threshold, and, second, showing that a single
threshold cannot simultaneously be safe for narrowband, and efficient
10
for wideband.
5
7.2.1 Difficulty in Using Thresholds 0 0 0 0
0
Method. This experiment uses one pair of SWIFT nodes at location
1 2 3 4 5 6 7 8 9 10
tx and rx in Fig. 8, and one pair of 802.11a nodes which is moved
among locations 1-10. At each location, we measure two quantities: Location
(a) Correct Bin Choice: We disable adaptive sensing on SWIFT and Figure 12: Adaptive Sensing is Robust: At each location, SWIFT
manually try all possible usable bin settings until we find the maximal finds the correct unusable bins, i.e. those that interfere with 802.11a.
set of usable bins that does not affect 802.11a throughput.
(b) Ideal Threshold: This is defined for each location as the highest
threshold that results in a bin choice which does not affect 802.11a in Results. Fig. 11 compares the number of wasted bins, i.e., bins that
that location. This is the threshold that is most efficient for wideband, the threshold unnecessarily marks as unusable by wideband, at each
while still being safe for narrowband. We record the time average location, against the corresponding 802.11a throughput, for both the
of the power SWIFT sees in each bin when 802.11a transmits, and median and minimum thresholds from Fig. 10. The median threshold
calculate the ideal threshold as the minimum power across all bins leads to a dramatic reduction in 802.11a throughput in locations 2, 3,
that must be left unused to ensure safe 802.11a operation. and 6, while simultaneously producing over 10 wasted wideband bins
Results. Fig. 10 shows the difficulty in choosing a single threshold in each of locations 8, 9, and 10. Bins are wasted in these locations
across locations: the ideal threshold varies by as much as 4.3× because the 802.11a nodes, being too far, are no longer affected by
in our testbed; furthermore, the thresholds do not correlate with wideband transmissions, but this threshold still causes many bins to
distance, because of the reflection and shadowing typical in an indoor be marked as unusable. Note that a threshold-based design can be
environment. both unsafe and inefficient in the same location. In particular, with
the median threshold it is unsafe in locations 2 and 6, but also wastes
a few bins in those same locations. This is because a blip in power in
7.2.2 No Single Threshold is Both Safe and Efficient any bin outside of those occupied by the narrowband device causes
In this section, we illustrate how a particular choice of threshold that bin to be wasted.
forces a compromise between safe narrowband operation and efficient A lower choice of threshold would increase the likelihood of safe
wideband performance across locations. narrowband operation at the cost of increased inefficiency. For ex-
Method. We use the same placement of wideband nodes as in §7.2.1. ample, using the minimum threshold among all measured locations
We consider two thresholds based on our experiments in §7.2.1 above, ensures safe 802.11a operation in all of these locations, but almost
setting the threshold to either the median, or the minimum of those in doubles the bandwidth wastage. In our example, in addition to wast-
Fig. 10. We then determine the set of bins that would be marked as ing bins in locations 7, 8, 9, and 10 where 802.11a is out of range,
usable for each threshold setting and location. We disable adaptive it also wastes bins in location 1. This wastage is because 802.11a
sensing in SWIFT, and at each location, manually set it to use the set transmissions leak significant power into bins adjacent to those it uses.
of bins resulting from the chosen threshold, and measure the 802.11a Additionally, this minimum threshold may be unsafe for locations
throughput. outside the measured set, or for a different 802.11a transmitter.
 30 1

Success Probability
802.11 Throughput (Mbps)
0.8
25
0.6
20
0.4
15 0.2
3) 802.11a
reacts 0
10
0 10 20 30 40 50 60
1) 802.11a Number of disagreeing bins
5 starts
Figure 14: Robustness to disagreement: The figure shows the
0
0 2 4 6 8 10 12 probability of a transmission succeeding as a function of the number
30 4) SWIFT retreats to of disagreeing bins. It shows that SWIFT is robust to as much as 40%
previous choice
25 disagreement between the set of transmitter and receiver bins.
Unusable Bins

20
23, and this returns the throughput of 802.11a to normal. As a result,
15 SWIFT stabilizes at a state that avoids bins 3 through 23, which is
10 the tightest bin selection that does not affect 802.11a.
5 2) SWIFT backs off
to conservative
bin choice 7.4 Dealing with Bin Disagreement
0
-5 We evaluate the impact of disagreement between communicating
0 2 4 6 8 10 12 pairs on SWIFT’s band consensus protocol.
Time (secs) Method. We place the wideband transmitter and receiver within a
few feet of each other so that they can communicate with each other
Figure 13: Responsiveness of Adaptive Sensing: The top graph with very low probability of channel bit errors. We do this to ensure
shows that 802.11a throughput is not hindered for longer than 0.5 that almost all bit errors are likely to be introduced purely due to
seconds by SWIFT. The bottom graph shows that, when 802.11a first disagreements. We initialize the transmitter and receiver to agree to
appears, SWIFT backs off to a conservative bin choice within 120 use the entire wide band, consisting of 100 bins.
ms, but quickly converges to a maximal set of safe bins. We then configure the adaptive sensing module to update the trans-
mitter with a new set of usable bins with a sequence of K consecutive
7.3 Adaptive Sensing bins marked as narrowband-occupied, to simulate the appearance of
a narrowband transmitter with a band of size K . Since the transmitter
In this section we show how the adaptive sensing algorithm allows
cannot use these bins whereas the receiver continues to expect data in
SWIFT to use a maximal set of bins with almost no impact on 802.11a,
them, the size of the disagreement between the nodes is K . We send
and hence is both safe and efficient.
a random coded sequence from transmitter to receiver using this dis-
Method. The setup is similar to the previous experiment, except that agreeing set of bins, check whether it is received correctly, and repeat
the SWIFT nodes now have adaptive sensing turned on. We run one this operation with a large number of random coded sequences for
experiment at each location, by first starting the SWIFT node, and increasing values of K . We declare a transmission to have succeeded
then starting the 802.11a transmission 5 seconds later. We record the if it is decoded correctly, and compute the probability of a successful
UsableBins setting on which SWIFT settles, and compare it with transmission for a disagreement of size K .
the correct bin setting for each location as determined in §7.2.1. Results. Fig. 14 shows that SWIFT’s band consensus works robustly
Results. Fig. 12 shows that SWIFT finds the exact set of unusable for a large range of disagreements. When K is small, the consensus
bins, i.e., bins that interfere with 802.11a, at all locations. Note scheme sees a very small number of errors which can be easily
further that SWIFT detects when 802.11a goes out of range, as in corrected. As K grows, the receiver sees a burst of errors in the
locations 7-10, and can reclaim all occupied bins. disagreeing bins, but the number of errors in any single code word
Fig. 13 shows the typical dynamics of adaptive sensing, using is limited because transmitted data bits are interleaved across the
results from an experiment with 802.11a at location 3. SWIFT con- frequency bins. This allows successful transmissions even when the
servatively backs away from bins used by 802.11a within 120 ms of fraction of disagreement is as large as 37% (37 of the 100 total bins).
802.11a commencing transmission. Additionally, within 4 seconds, it Such a large amount of disagreement is extremely unlikely, and hence
finds the ideal bin selection and then sticks with this selection. Over SWIFT’s low overhead handshake mechanism can almost always
60% of this time is a result of the communication overhead from our achieve band consensus. It is only when the extent of disagreement
prototype PCI driver, and can be mostly eliminated with an optimized becomes large (56 bins in our case) that SWIFT nodes will need to
implementation. reestablish connectivity using a sync packet.
Specifically, the bottom graph shows the SWIFT bin selections
over time. SWIFT starts out using all bins, (1) until it first detects the
7.5 Intermittent Narrowband TCP Web Downloads
802.11a transmissions. (2) At this point, SWIFT immediately backs
off using a conservative threshold, and avoids bins -2 through 28. This experiment evaluates SWIFT’s ability to adapt correctly to inter-
As it gathers more data, and determines that 802.11a is unaffected, mittent and bursty traffic patterns.
SWIFT decreases its set of unused bins gradually, till it begins avoid- Method. We model a typical home scenario, using an 802.11a node
ing only bins 4 through 22. (3) At this point, we see from the top that accesses the Internet by connecting to a Linksys wireless router.
graph that the throughput of 802.11a is affected for the first time. (4) We first start the SWIFT node, and at time t = 15 seconds, the
SWIFT immediately relaxes its bin selection to avoid bins 3 through 802.11a node begins periodic web downloads. For this experiment,
 3 choice. SWIFT converges on the right set of bins, and its throughput
802.11a Throughput (Mbps) stabilizes around t = 75 seconds. This throughput is lower than the
2.5 throughput that SWIFT achieved prior to the web downloads because
SWIFT is now avoiding bands that could affect 802.11a performance.
2
Throughout this process, SWIFT remains safe to 802.11a and does
1.5 not cause any noticeable impact on the TCP throughput.4

1
7.6 Network Results
0.5
Here, we show that SWIFT performs well even in a chaotic environ-
0 ment with multiple 802.11a devices, and multiple SWIFT nodes.
0 10 20 30 40 50 60 70 80 Method. In this experiment we use four wideband nodes and eight
802.11a nodes, creating six pairs of communicating nodes. We place
(a) 802.11a Throughput without SWIFT
the four 802.11a pairs at locations A-H, and the two wideband pairs
3 at the locations labeled tx/rx and tx’/rx’ in Fig. 8. We then measure
802.11a Throughput (Mbps)

the throughputs when running the network without any wideband

2.5
transmitters, with the wideband transmitters running NORM, and
2 with the wideband transmitters running SWIFT.
Results. Fig. 16(c) shows that, when NORM transmits simulta-
1.5 neously with 802.11a, it significantly reduces 802.11a throughput.
While the throughput reduction of 802.11a pairs at different locations
1
is different, all pairs are impacted, with an overall average loss in
0.5 throughput of around 50%.
Figs. 16(a) and (b) show the throughput of the four 802.11a pairs,
0 with and without SWIFT. In this case, both pairs of SWIFT nodes
0 10 20 30 40 50 60 70 80
move away from the bins occupied by the 802.11a nodes, allowing
(b) 802.11a Throughput with SWIFT all 802.11a pairs to have essentially the same performance as in the
absence of SWIFT. Additionally, Fig. 16(c) shows that by utilizing
200
SWIFT Throughput (Mbps)

all bins not occupied by 802.11a, the SWIFT nodes are each still able
to get reasonable throughputs of 30-100 Mbps in the face of 802.11a.
150 This result shows that SWIFT can deliver an operational wideband
network, while ensuring that it does not affect multiple competing
100 narrowband nodes.

50 8 Conclusion
0 This paper addresses the problem of coexistence between emerging
0 10 20 30 40 50 60 70 80 wideband networks and narrowband devices with which they share
Time (secs) the unlicensed bands. We show that overly conservative designs
(c) SWIFT Throughput with TCP that avoid interference by running below the noise floor needlessly
Figure 15: SWIFT reaction to TCP web downloads: (a) and (b) sacrifice the throughput and the range of the wideband radios. In
show that, even in the face of intermittent 802.11a traffic, SWIFT contrast, a design based on cognitive aggregation, which adapts
avoids affecting 802.11a transmissions, while (c) shows that it does its frequency bands and weaves together multiple non-contiguous
this while still achieving 90% of its original throughput. bands into one wireless link, can be as narrowband-friendly as the
conservative approaches, while achieving a significant increase in
operating range and throughput.
we download the home page from www.apple.com every 3 sec- Our results can be extended in multiple directions:
onds. We average the throughputs of the TCP downloads and SWIFT (a) Non-reactive narrowband devices: This paper addresses nar-
over 100ms intervals, and plot them as a function of time. rowband technologies that react to interference in their band. Of
Results. Fig. 15 shows that SWIFT adapts to intermittent and bursty course, not all devices react to interference. We envision that SWIFT
web traffic, without causing any performance impact on the narrow- can be extended to deal with such devices in one of two ways: ei-
band user. Notice that the narrowband traffic is indeed intermittent, ther by being configured to avoid known non-reactive bands if they
and that the TCP downloads are too short for narrowband to achieve are present, or by having adaptive sensing recognize a device as
a peak throughput higher than 2-3 Mbps, despite the fact that the non-reactive if all narrowband bins can be reclaimed without any
auto-rate algorithm is sustaining 48 or 54 Mbps in this case. identifiable reaction. In this case, SWIFT can fall back to a conser-
We see that SWIFT throughput drops as soon as the user begins her vative bin setting that avoids all bins with non-reactive narrowband
web download. This is because SWIFT falls back to a conservative power.
set of bins. SWIFT throughput then gradually increases as it tightens (b) Coexistence of multiple wideband protocols: SWIFT selec-
its set of bins. However, this process is slower than the example in tively avoids frequency bands used by narrowband devices, and
Fig. 13 because SWIFT only uses measurements in the vicinity of 4
The differences in TCP throughput with and without SWIFT are caused by varying
a narrowband transmission, as described in §5.1. It therefore needs queue lengths in the wired Internet. In particular, note that the variations in downloads
to wait for a longer time to acquire enough data points for each bin between the two graphs are no greater than the variations within any one graph.
 1 1
Fraction of 1-sec samples

Fraction of 1-sec samples

0.9 0.9
0.8 0.8
0.7 0.7
0.6 0.6
0.5 0.5
0.4 0.4
0.3 B-H 0.3 B-H
0.2 D-C 0.2 D-C
0.1 F-E 0.1 F-E
G-A G-A
0 0
0 2 4 6 8 10 12 0 2 4 6 8 10 12
802.11a Throughput (Mbps) 802.11a Throughput (Mbps)
(a) 802.11a Throughput alone (b) 802.11a Throughput with SWIFT
1 1
Fraction of 1-sec samples

Fraction of 1-sec samples

0.8 0.8

0.6 0.6

0.4 0.4
B-H
0.2 D-C 0.2
F-E SWIFT Pair 1
G-A SWIFT Pair 2
0 0
0 2 4 6 8 10 12 0 20 40 60 80 100 120
802.11a Throughput (Mbps) SWIFT Throughput (Mbps)
(c) 802.11a Throughput with NORM (d) SWIFT concurrent with 802.11a in (b)
Figure 16: Throughputs in a Network: (a) and (b) show the throughputs of the four 802.11a pairs, with and without SWIFT. SWIFT has no
impact on 802.11a, while, still getting good throughput as seen in (d). In contrast, (c) shows that non-adaptive wideband transmitters reduce
802.11a throughput by around 50%.

shares the spectrum with other cooperating wideband devices us- [14] V. Brik, E. Rozner, S. Banerjee, and P. Bahl. DSAP: A Protocol for Coordinated Spectrum
Access. In IEEE DySPAN 2005.
ing the SWIFT protocol. However, the future may bring a variety
[15] M. M. Buddhikot, P. Kolodzy, S. Miller, K. Ryan, and J. Evans. DIMSUMNet: New Directions
of wideband protocols. These systems need to find a way to share in Wireless Networking Using Coordinated Dynamic Spectrum Access. 2005.
spectrum among different wideband technologies even when they do [16] D. Cabric, S. M. Mishra, D. Willkomm, R. Brodersen, and A. Wolisz. A Cognitive Radio
Approach for Usage of Virtual Unlicensed Spectrum. In 14th IST Wireless Comms. Summit.
not use the same protocol. [17] G. Casella and R. L. Berger. Statistical Inference. Duxbury, 2nd edition, 2002.
(c) Dynamic Range: Like other techniques that allow a node to [18] M. S.-W. Chen and R. Brodersen. A Subsampling UWB Impulse Radio Architecture Utilizing
receive multiple concurrent signals [33], SWIFT’s nodes deal with a Analytic Signaling. IEICE Transactions on Electronics, Vol, E88-C, 2005.
[19] S. S. Company. Spectrum occupancy measurement, 2007. http://www.sharedspectrum.
wide range of signal powers and hence their performance improves com/measurements/.
with a wider dynamic range of the system. [20] F. Edalat. Real-time Sub-carrier Adaptive Modulation and Coding in Wideband OFDM Wireless
Systems. PhD thesis, Massachusetts Institute of Technology, 2008.
Acknowledgments: We thank Manish Bharadwaj for several insight- [21] S. C. Ergen. ZigBee/IEEE 802.15.4 summary, 2004.
[22] M. S. Gast. 802.11 Wireless Networks. O’Reilly, 2nd edition, 2005.
ful discussions on the design of SWIFT, and Szymon Chachulski,
[23] J. Heiskala and J. Terry. OFDM Wireless LANs: A Theoretical and Practical Guide. Sams
Micah Brodsky, and Rob Beverly for their helpful comments on the Publishing, 2001.
paper. This work is funded by DARPA ITMANET and Quanta. [24] V. Jacobson. Congestion avoidance and control. In ACM SIGCOMM ’88, Stanford, CA, 1988.
[25] L. Ma, X. Han, and C.-C. Shen. Dynamic Open Spectrum Sharing MAC Protocol for Wireless
Ad Hoc Networks. In IEEE DySPAN 2005.
[26] N. Matalon. An Implementation of a 5.25 GHz Transceiver for High Data Rate Wireless Appli-
References cations. MS thesis, MIT, EECS, July 2005.
[27] M. McHenry. Frequency Agile Spectrum Access Technologies. Presentation to 2003 FCC
[1] Cutting the Cord to Flat-screen TVs. CNN Jan 2008. Workshop on Cognitive Radios.
[2] FCC Slides for UWB Spectral Limits. http://sss-mag.com/uwbslides.html. [28] S. Mishra, S. ten Brink, R. Madadevappa, and R. Brodersen. Detect and Avoid: An Ultra-
[3] IEEE 802.22 WG. www.ieee802.org/22/. Wideband/WiMax Coexistence Mechanism. IEEE Communications Magazine, June 2007.
[4] Impact of devices using UWB technology on radiocommunication services. ITU R SM.1757. [29] M. Mittelbacht, C. Mullert, D. Fergert, and A. Fingert. Study of Coexistence Between UWB
and Narrowband Cellular Systems. In UWB Systems, 2004.
[5] UWB - Intel Standards. http://intel.com/standards/case/case_uwb.htm.
[30] J. Proakis and M. Salehi. Digital Communications. McGraw-Hill, 5th edition, 2007.
[6] Xilinx. http://xilinx.com/products/design_resources/design_tool/.
[31] H. Rahul, N. Kushman, D. Katabi, C. Sodini, and F. Edalat. SWIFT: A Narrowband-Friendly
[7] Local and Metropolitan Area Networks Specific Requirements Part 11: Wireless LAN Medium
Cognitive Wideband Network. Technical Report MIT-CSAIL-TR-2008, MIT, 2008.
Access Control (MAC) and Physical Layer (PHY), 1999.
[32] R. Tandra and A. Sahai. SNR walls for signal detection. In IEEE Journal on Special Topics in
[8] Local and Metropolitan Area Networks Specific Requirements Part 11: Wireless LAN Medium
Signal Processing, Feb. 2008.
Access Control (MAC) and Physical Layer (PHY), 2003.
[33] D. Tse and P. Vishwanath. Fundamentals of Wireless Communications. Cambridge University
[9] Enabling high-speed wireless personal area networks, 2005. Intel White Paper.
Press, 2005.
[10] I. F. Akyildiz, W. Y. Lee, M. Vuran, and S. Mohanty. NeXt Generation Dynamic Spectrum
[34] C. Wilmot. Intel demonstrates fast new Ultrawideband WPAN at IDF Taiwan, 2006.
Access Cognitive Radio Wireless Networks: A survey.
[35] J. Zhao, H. Zheng, and G. Yang. Distributed Coordination in Dynamic Spectrum Allocation
[11] J. G. Andrews, A. Ghosh, and R. Muhamed. Fundamentals of WiMAX: Understanding Broad-
Networks. In IEEE DySPAN 2005.
band Wireless Networking. PrenticeHall, 2007.
[36] Q. Zhao, L. Tong, and A. Swami. Decentralized cognitive MAC for dynamic spectrum access.
[12] P. Bahl, R. Chandra, P. A. Chou, J. I. Ferrell, T. Moscibroda, S. Narlanka, and Y. Wu. KNOWS:
In IEEE DySPAN 2005.
Kognitiv Networking Over White Spaces. In IEEE DySPAN 2007.
[37] ZigBee Alliance. ZigBee and Wireless Radio Frequency Coexistence, June 2007. White Paper.
[13] H. Bölcskei. Principles of MIMO-OFDM wireless systems. 2004.

View publication stats