HARNESSING LOW-PASS FILTER DEFECTS FOR IMPROVING

WIRELESS LINK PERFORMANCE: MEASUREMENTS AND

APPLICATIONS

By

Alireza Ameli Renani

A DISSERTATION

Submitted to

Michigan State University

in partial fulﬁllment of the requirements

for the degree of

Computer Science - Doctor of Philosophy

2018

ABSTRACT

Harnessing Low-Pass Filter Defects for Improving Wireless Link Performance:

Measurements and Applications

By

Alireza Ameli Renani

The design trade-oﬀs of transceiver hardware are crucial to the performance of wireless sys-

tems. The eﬀect of such trade-oﬀs on individual analog and digital components are vigorously

studied, but their systemic impacts beyond component-level remain largely unexplored. In

this dissertation, we present an in-depth study to characterize the surprisingly notable sys-

temic impacts of low-pass ﬁlter design, which is a small yet indispensable component used

for shaping spectrum and rejecting interference.

Using a bottom-up approach, we examine how signal-level distortions caused by the

trade-oﬀs of low-pass ﬁlter design propagate to the upper-layers of wireless communica-

tion, reshaping bit error patterns and degrading link performance of today’s 802.11 systems.

Moreover, we propose a novel unequal error protection algorithm that harnesses low-pass ﬁl-

ter defects for improving wireless LAN throughput, particularly to be used in forward error

correction, channel coding, and applications such as video streaming. Lastly, we conduct

experiments to evaluate the unequal error protection algorithm in video streaming, and we

present substantial enhancements of video quality in mobile environments.

Dedicated to my parents, Roya and Ahmad;

and to my wife, Elaheh.

iii

ACKNOWLEDGEMENTS

First of all, I want to thank my adviser, Dr. Guoliang Xing. You helped me become an

independent researcher. Thank you for taking a chance on me, I truly appreciate all your

support.

To Dr. Abdol-Hossein Esfahanian, I cannot thank you enough for encouraging me to

apply for PhD here at MSU, and for all your help and support during this journey. It has

been a pleasure to work on projects under your supervision. You have been there whenever

I needed any help or guidance. I am truly grateful and honored.

I want to extend my heartfelt thanks to Dr. Richard Enbody. I could not ask for

a better mentor. I learned how to be passionate in teaching by observing how you teach,

support and lead. I have learned so much from you I cannot express in words. Your oﬃce

door is always literally and metaphorically open to everyone, and I stopped by your oﬃce

countless number of times. I am truly honored to have a chance to teach beside you and call

you my colleague. Thank you so much for all your mentorship and support.

I want to thank my committee member, Dr. Matt Mutka. I appreciate all your help

and support which made my PhD completion much more achievable. I further want to extend

my thanks to my committee member, Dr. Hayder Radha. I appreciate your guidance and

insights whenever I needed them. I also Want to thank Dr. Jun Huang for his collaborations

in my projects.

To my parents, Roya and Ahmad, thank you for always encouraging me to follow my

intellectual curiosity my entire life. I have achieved so many personal and professional goals

thanks to your never ending encouragements, patience and support. Along with my brother,

iv

Mohammd Javad and my sister, Zahra, I want to thank you all for always being on my side

through all the ups and downs in my life. I am truly blessed to have a wonderful family like

you.

To all my friends, thank you all so much for always standing by me and making my

journey more tolerable. I wish I could name you all. I wish you the best of luck wherever

you are.

To my wonderful wife, Elaheh, thank you so much for being with me this entire

journey. Meeting you was the best thing that happened in my life. Thank you for being

such an amazing partner and standing by me no matter what life threw at us. I am looking

forward to the next chapter of our lives together.

v

TABLE OF CONTENTS

LIST OF TABLES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viii

LIST OF FIGURES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

ix

LIST OF ALGORITHMS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiii

Chapter 1

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.1 Background and Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2 Problem Statement and Goals . . . . . . . . . . . . . . . . . . . . . . . . . .

Chapter 2 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1 Cross-Layer Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2 Error Estimation and Recovery . . . . . . . . . . . . . . . . . . . . . . . . .
2.3 Video Streaming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.4 Vehicular Communication . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Chapter 3 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.1
802.11 Primer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 Low-Pass Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Chapter 4 The Impacts of LPF Defects Under the Microscope . . . . . .
4.1 Signal-Level Characterization . . . . . . . . . . . . . . . . . . . . . . . . . .
4.1.1 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.1.2 Measuring Frequency Roll-Oﬀ . . . . . . . . . . . . . . . . . . . . . .
Signiﬁcance in Practical Environments
4.1.3
. . . . . . . . . . . . . . . . .
4.1.4
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Impacts on Upper Layers . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2.1 From Signal-Level to Bit-Level . . . . . . . . . . . . . . . . . . . . . .
4.2.2 Measuring Bit Error Patterns . . . . . . . . . . . . . . . . . . . . . .
4.2.3
. . . . . . . . . . . . . . . . . . . . . . .

Simulation-Based Diagnosis

4.2

Chapter 5 Application in Video Streaming . . . . . . . . . . . . . . . . . . .
5.1 The Unequal Error Protection Algorithm . . . . . . . . . . . . . . . . . . . .
5.1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.1.2 Basic Idea . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.1.3 Algorithm Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2
Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.3 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.3.1
Setup and Conﬁguration . . . . . . . . . . . . . . . . . . . . . . . . .
5.3.2 Baseline and Metric . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.3.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

vi

1
1
2

5
5
5
6
7

10
10
11

14
14
15
18
20
23
24
24
26
30

34
34
35
36
37
40
41
42
43
43

Chapter 6 Application in Vehicular Communication . . . . . . . . . . . . .
6.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.2 Harnessing Hardware Defects in Vehicular Communication . . . . . . . . . .
6.3 Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.3.1 Hardware Defect Modeling . . . . . . . . . . . . . . . . . . . . . . . .
6.3.2 Harnessing Hardware Defects at PHY . . . . . . . . . . . . . . . . . .
6.3.3 Exploiting BER Pattern at Link Layer . . . . . . . . . . . . . . . . .
6.4 Suggested Evaluation Methodology . . . . . . . . . . . . . . . . . . . . . . .

63
63
65
66
66
67
67
68

Chapter 7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

72

REFERENCES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

73

vii

LIST OF TABLES

Table 4.1: 802.11g OFDM modulations, coding and data rates [31] [61].

. . . . . . .

30

Table 5.1: PSNR to MOS conversion [41] [51]. . . . . . . . . . . . . . . . . . . . . . .

43

Table 5.2: Average PSNR and SSIM of all the experiments.

. . . . . . . . . . . . . .

48

Table 5.3: Average byte errors in I, P and B frames in standard and UEP streamings
. . . . . . . . . . . . . . . . . . . . . . . . . . . .

without retransmission.

Table 5.4: Average byte errors in I, P and B frames in standard and UEP streamings
. . . . . . . . . . . . . . . . . . . . . . . . . . . .

with 3 retransmissions.

62

62

viii

LIST OF FIGURES

Figure 1.1: Normalized bit error frequency for transmission rate 54Mbps.

. . . . . .

2

Figure 3.1: 802.11 architecture at a high level. Figure illustrates scrambling, coding
and interleaving in the architecture. . . . . . . . . . . . . . . . . . . . . .

11

Figure 3.2: Transmit spectrum mask for 20MHz transmission [7].

. . . . . . . . . . .

12

Figure 3.3:

Impulse response of Rect function, Sinc(x). . . . . . . . . . . . . . . . . .

13

Figure 3.4: Eﬀects of ﬁlter order on roll-oﬀ: Higher order ﬁlters have steeper transi-
. . . . . . . . . . . . . . . . . . . . .

tions and smaller transition bands.

Figure 4.1: Architecture of the receiver front-end, consisting of a USRP2 and an SBX
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

daughterboard.

Figure 4.2: Magnitude response of CIC ﬁlter for the USRP2.
. . . . . . . . . . . . .
Figure 4.3: Setup of over-cable measurements using an USRP and a 2ˆ2 MIMO
AR9285 chip. The AR9285 is connected to an Agilent 8494B attenua-
tor through an RCA DH24SP 2-way signal splitter, and the signal was
measured from the main antenna port.
. . . . . . . . . . . . . . . . . . .

Figure 4.4: Transmitter power spectral densities of diﬀerent 802.11 chipsets measured
before and after compensation for CIC roll-oﬀ. . . . . . . . . . . . . . . .

Figure 4.5: Power losses caused by transmitter LPF on border subcarriers (subcarriers
-26 and +26) and center subcarriers (subcarriers -1 and +1). Error bars
show the minimum and maximum measurement results.
. . . . . . . . .

Figure 4.6: Projected power losses on subcarrier 26 for diﬀerent transmitter and re-
ceiver pairs. The results were computed based on the measurement shown
in Fig. 4.5.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

13

16

17

19

20

21

22

Figure 4.7: Transmitter spectral densities measured at three randomly selected static
links within a single oﬃce building. The LPF-induced roll-oﬀ became
insigniﬁcant as a result of the severe distortion caused by multipath fading. 23

Figure 4.8: CDFs of power losses measured in the presence of multipath fading. The
power losses at channel center (i.e., subcarrier +1 and -1) and channel
border (i.e., subcarrier +26 and -26) are similar, as the eﬀect of LPF
roll-oﬀ is overwhelmed by multipath fading.
. . . . . . . . . . . . . . . .

24

ix

Figure 4.9: The average PSD measured on an 802.11 receiver moving at walking speed.
LPF-induced roll-oﬀ becomes increasingly pronounced as the number of
received packets increases. Roll-oﬀ is the most signiﬁcant error source in
the average PSD measured using 500 packets.
. . . . . . . . . . . . . . .

Figure 4.10: The average of power losses measured at channel border (i.e., subcarrier
+26 and -26) and channel center (i.e., subcarrier +1 and -1) as the number
of received packets increases. . . . . . . . . . . . . . . . . . . . . . . . . .

25

26

Figure 4.11: Bits close to the border suﬀer from in-band signal loss.

. . . . . . . . . .

27

Figure 4.12: Byte error rates for transmission rate of 54Mbps. The vertical axis repre-
sents the normalized bit error rate and the horizontal axis represents the
byte position within the payload.
. . . . . . . . . . . . . . . . . . . . . .

Figure 4.13: Byte error rates for transmission rate of 48Mbps. The vertical axis repre-
sents the normalized bit error rate and the horizontal axis represents the
byte position within the payload.
. . . . . . . . . . . . . . . . . . . . . .

Figure 4.14: The autocorrelation of error rates over bit positions under diﬀerent trans-
mission rates. The period of the error pattern equals to the number of
bits carried by one OFDM symbol. The transmission rates are 48Mbps
. . . . . . . . . . . .
and 54Mbps for left and right ﬁgures, respectively.

Figure 4.15: Filter simulation results illustrating the eﬀect of cut-oﬀ frequency and
ﬁlter order on packet success rate. . . . . . . . . . . . . . . . . . . . . . .

Figure 4.16: Heat map of ﬁlter simulation results illustrating the eﬀect of cut-oﬀ fre-
quency and ﬁlter order on packet success rate. . . . . . . . . . . . . . . .

Figure 4.17: Bit error rates measured using a) WiFi link of two AR9285 chips; b) simu-
lation of a 6-tap ﬁlter with cut-oﬀ frequency of 5 subcarriers; c) simulation
of a 32-tap ﬁlter with cut-oﬀ frequency of 2 subcarriers. . . . . . . . . . .

Figure 5.1: Standard video streaming is used to transmit a GOP with one I-frame,
. . . . . . . .

one P-frame and two B-frames using ﬁve wireless packets.

Figure 5.2: UEP algorithm is used to transmit a GOP with one I-frame, one P-frame
and two B-frames using ﬁve wireless packets. . . . . . . . . . . . . . . . .

Figure 5.3: UEP algorithm maps higher priority data to bit positions that experience
. . . . . . . . . . . . . . . . . . .

lowest error rates in a wireless packet.

28

28

29

32

32

33

39

39

40

x

Figure 5.4: A video frame of akiyo.mp4 recorded from standard streaming. Many
parts of the frame are completely corrupted. With PSNR of 12 it falls
into the lowest MOS category. . . . . . . . . . . . . . . . . . . . . . . . .

Figure 5.5: A video frame of akiyo.mp4 recorded from UEP streaming which shows
major improvements in quality of the video. With PSNR of 30, it is in
the high range of fair category in MOS. . . . . . . . . . . . . . . . . . . .

Figure 5.6: PSNR of all 300 frames in akiyo.mp4 streamed using UEP and standard
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

streamings.

Figure 5.7: MOS of all the frames in akiyo.mp4 streamed using UEP and standard
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

streamings.

Figure 5.8: PSNR of all 300 frames in akiyo.mp4 streamed using UEP and standard
. . . . . . . . . . . . . . . . . . . .

streamings in experiment number 1.

Figure 5.9: MOS of all the frames in akiyo.mp4 streamed using UEP and standard
. . . . . . . . . . . . . . . . . . . .

streamings in experiment number 1.

Figure 5.10: PSNR of all 300 frames in akiyo.mp4 streamed using UEP and standard
. . . . . . . . . . . . . . . . . . . .

streamings in experiment number 2.

Figure 5.11: MOS of all the frames in akiyo.mp4 streamed using UEP and standard
. . . . . . . . . . . . . . . . . . . .

streamings in experiment number 2.

Figure 5.12: PSNR of all 300 frames in akiyo.mp4 streamed using UEP and standard
. . . . . . . . . . . . . . . . . . . .

streamings in experiment number 3.

Figure 5.13: MOS of all the frames in akiyo.mp4 streamed using UEP and standard
. . . . . . . . . . . . . . . . . . . .

streamings in experiment number 3.

Figure 5.14: PSNR of all 300 frames in akiyo.mp4 streamed using UEP and standard
. . . . . . . . . . . . . . . . . . . .

streamings in experiment number 4.

Figure 5.15: MOS of all the frames in akiyo.mp4 streamed using UEP and standard
. . . . . . . . . . . . . . . . . . . .

streamings in experiment number 4.

Figure 5.16: PSNR of all 300 frames in akiyo.mp4 streamed using UEP and standard
. . . . . . . . . . . . . . . . . . . .

streamings in experiment number 5.

Figure 5.17: MOS of all the frames in akiyo.mp4 streamed using UEP and standard
. . . . . . . . . . . . . . . . . . . .

streamings in experiment number 5.

44

45

46

47

49

50

52

52

53

53

54

54

55

55

xi

Figure 5.18: PSNR of all 300 frames in akiyo.mp4 streamed using UEP and standard
. . . . . . . . . . . . . . . . . . . .

streamings in experiment number 6.

Figure 5.19: MOS of all the frames in akiyo.mp4 streamed using UEP and standard
. . . . . . . . . . . . . . . . . . . .

streamings in experiment number 6.

Figure 5.20: PSNR of all 300 frames in akiyo.mp4 streamed using UEP and standard
. . . . . . . . . . . . . . . . . . . .

streamings in experiment number 7.

Figure 5.21: MOS of all the frames in akiyo.mp4 streamed using UEP and standard
. . . . . . . . . . . . . . . . . . . .

streamings in experiment number 7.

Figure 5.22: PSNR of all 300 frames in akiyo.mp4 streamed using UEP and standard
. . . . . . . . . . . . . . . . . . . .

streamings in experiment number 8.

Figure 5.23: MOS of all the frames in akiyo.mp4 streamed using UEP and standard
. . . . . . . . . . . . . . . . . . . .

streamings in experiment number 8.

Figure 5.24: PSNR of all 300 frames in akiyo.mp4 streamed using UEP and standard
. . . . . . . . . . . . . . . . . . . .

streamings in experiment number 9.

Figure 5.25: MOS of all the frames in akiyo.mp4 streamed using UEP and standard
. . . . . . . . . . . . . . . . . . . .

streamings in experiment number 9.

Figure 5.26: PSNR of all 300 frames in akiyo.mp4 streamed using UEP and standard
streamings in experiment number 10. . . . . . . . . . . . . . . . . . . . .

Figure 5.27: MOS of all the frames in akiyo.mp4 streamed using UEP and standard
streamings in experiment number 10. . . . . . . . . . . . . . . . . . . . .

56

56

57

57

58

58

59

59

60

60

Figure 6.1: The architecture of the system.

. . . . . . . . . . . . . . . . . . . . . . .

66

Figure 6.2: Setup of over-cable measurements using an USRP2 and Cohda MK5 OBU.
The MK5 is connected to a Mini-Circuits VAT-20+ attenuator, and the
signal was measured from the antenna port 1.
. . . . . . . . . . . . . . .

69

Figure 6.3: Setup of measurements using two MK5 units over a stationary link. The

antenna is MGW-303 which includes two WiFi antennas and a GPS antenna. 70

xii

LIST OF ALGORITHMS

1

Pseudocode of the UEP algorithm.

. . . . . . . . . . . . . . . . . . . . . . .

51

xiii

Chapter 1

Introduction

1.1 Background and Motivation

With the rapid advance of physical layer (PHY) technology, the implementation of

802.11 transceiver hardware is becoming increasingly complex. From 802.11g-2003 [7] to

802.11n-2009 [8], the number of transistors required to implement the transceiver increased

12.5 times in just six years [20] [65], which is 1.56 times faster than the Moore’s rate. This

trend is expected to continue in 802.11ac and the upcoming 802.11ad [9] [10]. To cope with

this trend, a key trade-oﬀ in transceiver design is to balance the performance and cost of

hardware implementation. As an example of such trade-oﬀs, implementing a ﬂash analog-
to-digital converter (ADC) of n-bit resolution requires 2n ´ 1 comparators. The better the

precision of conversion, the larger the package size and the higher the power consumption.

Similarly, the gain and stability of a low-noise ampliﬁer (LNA) can be improved by increasing

current density, which is however at the expense of higher current draw [29].

The design trade-oﬀs in transceiver hardware are crucial to the performance of wire-

less systems. However, although the eﬀects of such trade-oﬀs on individual analog and

digital components are well-studied, their systemic impacts beyond component-level remain

largely unexplored. The problem is particularly complex in commercial wireless systems,

where consumers and researchers are not privy to the proprietary design choices, resulting in

1

Figure 1.1 Normalized bit error frequency for transmission rate 54Mbps.

hardware-related communication and performance issues at upper-layers that are diﬃcult to

diagnose and control [34] [35] [50] [64]. For example, recent work [34] observed that oﬀ-the-

shelf 802.11 devices suﬀer from performance degradation due to a twisted bit error pattern

shown in Fig. 1.1, where bits transmitted at certain positions of a frame experience signiﬁ-

cantly higher error rates. This pattern persists even in over-cable communication, implying

that the cause is rooted in transceiver hardware. Although a number of protocols have been

developed to optimize communication performance based on this pattern [37] [57], the root

cause of such error pattern remains unidentiﬁed.

The contents of this chapter were adapted from a published conference paper 1.

1.2 Problem Statement and Goals

In order to precisely understand the performance of today’s wireless systems, we

present an in-depth study to characterize the surprisingly notable systemic impacts of hard-

ware defects caused by component-level design trade-oﬀs. To this end, we take Low-Pass

1“Harnessing hardware defects for improving wireless link performance: Measurements and

applications”. Alireza Ameli Renani, Jun Huang, Guoliang Xing, Abdol-Hossein Esfahanian.

INFOCOM 2017 - IEEE Conference on Computer Communications.

2

02004006008001000Bit Position0.00060.00080.00100.00120.00140.00160.00180.00200.0022Normalized Bit Error RateFilter (LPF) – a small yet indispensable component used for shaping spectrum and rejecting

interference – as an example, and study the impacts of its defects on the performance of

oﬀ-the-shelf 802.11 systems. Our contribution is three-fold:

1. Measurement of signal-level distortions caused by LPF defects Signal-level distortions

play a signiﬁcant role in overall wireless system performance.

It is thus crucial to

measure and understand such distortions caused by hardware components like LPF. A

basic trade-oﬀ in LPF design is to balance the steepness of frequency roll-oﬀ with im-

plementation complexity. A slow roll-oﬀ may distort wanted signal components in the

passband, while a steep roll-oﬀ requires higher implementation costs. To understand

the eﬀects of such trade-oﬀ in commercial wireless systems, we present microscopic

measurements and models to characterize the signal-level distortions caused by LPF

roll-oﬀ, and evaluate its signiﬁcance in practical environments in the presence of mul-

tipath fading and mobility.

2. Characterization of the impacts of LPF defects on link reliability Using a bottom-up

approach, we examine how signal-level defects caused by LPF roll-oﬀ propagate to the

upper-layers of wireless communication, reshaping bit error patterns and degrading

link performance. To this end, we build standard-compliant emulation tools on a

software-deﬁned radio to measure and diagnose the impacts of LPF, and quantify the

degradation of link reliability caused by LPF defects.

3. Harnessing LPF defects for mobile video streaming In addition to measuring and mod-

eling the eﬀects of LPF, we propose algorithms that harness LPF defects for improving

mobile video streaming, which entails signiﬁcant challenge due to fast varying channel

conditions. Our approach is driven by the insight that distortions caused by hard-

3

ware defects are invariant of wireless channel. As a result, although the instantaneous

performance of video streaming is highly dependent on channel condition, hardware

defects typically result in error patterns that are stable and predictable on mobile

links, after the distortion introduced by wireless channel is averaged out after a short-

period of movement. Based on this observation, we design an unequal error protection

(UEP) algorithm that reorders video frame bytes based on the error pattern caused

by LPF defects, which substantially enhances video streaming performance in mobile

environments.

4

Chapter 2

Related Work

2.1 Cross-Layer Design

Numerous cross-layer protocols have been developed for various wireless networks.

Channel information is used in [55] to provide unequal error protection for upper-layer traf-

ﬁc ﬂows. The fact that decoded erroneous symbols are still a good “approximation” of the

original transmitted symbol is the foundation in designing a UEP algorithm. [23] leverages

CSI to recover partially corrupted FEC groups and facilitate FEC decoding. EﬀSNR [23], Ac-

cuRate [56] and SoftRate [59] exploit various PHY hints to facilitate rate adaptation. These

cross-layer designs exploit PHY information at the link layer or above to improve wireless

performance. TiM [60] or time-line modulation, adds time dimension into amplitude-phase

scheme for a three-dimensional modulation scheme to ﬁll the gap of not having a one-to-one

mapping from channel conditions to modulation schemes.

2.2 Error Estimation and Recovery

Maranello [36] and PPR [40] reduce unnecessary retransmissions by only retrans-

mitting bit blocks that contain errors. ZipTx [45] sends parity bits to correct bit errors.

SOFT [62] allows multiple receivers to cooperatively repair corrupted packets. Bit error rate

5

estimation is central to various wireless protocols. Previous solutions either use error esti-

mation codes (EEC) [28], or PHY hints for bit error rate (BER) estimation. However, EEC

requires transmitter cooperation for code embedding, while soft-values are not available on

many production wireless NICs. Han et al. [34][35] identiﬁed the bit error pattern using

experimental results and hypothesize their cause for upper-layer applications.

2.3 Video Streaming

There is a large body of work on video coding and video streaming to increase the

throughput, quality and performance of the video streaming. SoftCast [39] broadcasts video

without specifying a data rate and each receiver obtains a video quality depending on its

channel quality. Recitation [44] employs CSI to predict error-prone bit positions and expose

them to upper layers. FlexCast [19] provides a video codec to increase video streaming per-

formance. piStream [63] estimates the bandwidth by monitoring the physical layer resource

allocation and enhances adaptive video streaming over LTE by utilizing an agile video rate

adaptation. A-HDAVT [66] is an adaptive hybrid digital-analog video transmission scheme

that is designed for video streaming in mobile networks. The study combats channel fading

and improves the system performance unlike most previous studies that improve the system

only in unrealistic Gaussian channels.

6

2.4 Vehicular Communication

VANETs are about to be deployed on a large scale, hence recent papers studied

diﬀerent aspects of it, from modeling and simulations to ﬁeld studies and physical layer

analysis. Physical layer eﬀects are very hard to model and there is a need to understand the

performance of IEEE802.11p receivers.

Bloessl et al. took the ﬁrst steps toward an open source 802.11p stack by implement-

ing a complete OFDM transceiver and extending it to a simulation and experimentation

framework for 802.11p [25] [26]. They validated the implementation by testing it using mul-

tiple wiﬁ devices and two 802.11p prototypes, Cohda MK2 and Unex DCMA-86P2. They

further conducted a Field Operational Test (FOT) using a modiﬁed 802.11a chipset as the

transmitter and an Software Deﬁned Radio (SDR) as the receiver [24]. They collected traces

in an hour long test over 56 km in city, rural, freeway and highway environments. They

further evaluated diﬀerent receive algorithms on the collected data. Another SDR imple-

mentation of IEEE 802.11p is presented in [47]. The authors focus on power eﬃciency of the

implementation which is prohibitive in mobile environments.

Field trials for studying DSRC in VANETs can be very expensive specially when the

number of vehicles involved is high. Therefore, many studies used network or physical layer

simulations. The simulators often abstract details of the other layers or even do not consider

them at all, which is problematic. Recent study proposed a solution by integrating IEEE

802.11p physical layer simulator with NS-3 which is a discrete-event network simulator[49].

Bernado et al. modeled and simulated the physical layer of 802.11p and studied the

eﬀect of many factors on the channel estimation performance such as strength of diﬀuse

components, Line-Of-Sight (LOS) and SNR [22]. Sassi et al. modeled and simulated IEEE

7

802.11p PHY and evaluated the OFDM transmission performance by studying the aﬀect of

various parameters on Bit Error Rate (BER) versus SNR[54].

Qualcomm Research [12] has introduced Cellular V2X (C-V2X) technology which is

designed to connect vehicles to vehicles, infrastructure, pedestrians, network and basically

everything (V2X). It can coexist with 802.11p and its trials start in 2017 with commercial

solutions.

IEEE 802.11p speciﬁes four Spectrum Emission Masks (SEMs) which introduces a

challenge for ﬁltering signals in 802.11p to meet the standard. Recent work investigated

multiple techniques and proposed a method to relax the constraints of radio frequency front-

end such that implementation of 802.11p system would be possible by using 802.11a hard-

ware [52].

Experimental measurements and studies of WAVE started long before IEEE 802.11p

was published in 2010. Kukshya et al. conducted various experiments using modiﬁed 802.11a

chipsets [43]. They considered line-of-sight scenarios with no vehicle traﬃc, moderate vehicle

traﬃc and high vehicle density in parking lot, street and freeway environments. They studied

the eﬀect of SNR on BER and also the eﬀect of distance on SNR of the received packets.

Based on small diﬀerences between IEEE 802.11p and other IEEE 802.11 standards,

[58] presents an approximation of 802.11p standard using commercial oﬀ-the-shelf 802.11a

hardware and some software adjustments. Using the proposed implementation makes it pos-

sible to use cheap 802.11a hardware compared to rather expensive 802.11p prototypes for

VANET research. It can also turn available WLAN testbeds into valuable VANET research

testbeds. The most important drawback of this work and many other similar implemen-

tations is that not all parameters of 802.11p can be modiﬁed on a 802.11a hardware. For

example, they used 20 MHz channels instead of 10 MHz channels, which means parameters

8

such as symbol duration, guard time and subcarrier spacing do not represent a real 802.11p

hardware.

Most of the prior work in this ﬁeld focus on physical layer information to increase

the communication performance while we study and utilize the hardware defects to create

our state of the art unequal error protection algorithm which can improve communication

performance in any scenario that data is divided into diﬀerent levels of importance such as

error estimation and recovery, video streaming and vehicular communication.

9

Chapter 3

Background

In this chapter we present a brief background on 802.11 architecture, OFDM basics

and Low Pass Filter which are the subjects of our study.

3.1

802.11 Primer

Fig. 3.1 shows the high level architecture of 802.11. The transmitter ﬁrst scrambles

the original bit-stream and randomizes the data bits. Then the data bits are coded using

convolutional encoding. Orthogonal Frequency Division Multiplexing (OFDM) divides the

20 MHz 802.11g channel bandwidth into 64 312.5 KHz subcarriers. 802.11g uses 4 pilot

subcarriers, 12 null subcarriers and the remaining 48 subcarriers for data. The center or

“DC” subcarrier is not used. Out of the remaining 11 subcarriers, 6 of them are in the

beginning of the channel and 5 of them are at the end. The transmitter interleaves and

maps the coded bits to subcarriers and also modulates them into symbols. The combined

signal of all subcarriers is called OFDM symbol. Each received OFDM symbol undergoes

the reverse operations to recover the original data.

According to Part 11 of IEEE Standard for Wireless LAN MAC and PHY Speciﬁca-

tion, the transmitted spectrum using 20 MHz channel spacing should fall within the spectral

mask described below.

The following items and Fig. 3.2 present the spectral mask that the signal should fall

10

Figure 3.1 802.11 architecture at a high level. Figure illustrates scrambling, coding and
interleaving in the architecture.

within. The units are in dBr which is dB relative to the maximum spectral density of the

signal. The transmitted spectrum should have a 0 dBr bandwidth not exceeding 18 MHz

centered at fc , the channel center. -20 dBr at 11 MHz frequency oﬀset. -28 dBr at 20 MHz

frequency oﬀset. The maximum of 40 dBr and 53 dBm/MHz at 30 MHz frequency oﬀset

and above.

3.2 Low-Pass Filter

Low-pass ﬁlter is widely used in wireless systems for reshaping spectrum and rejecting

interference. An ideal low-pass ﬁlter passes signals with frequencies lower than a speciﬁed

cutoﬀ frequency, and eliminates other signals with higher frequencies. The frequency re-

sponse of the ideal LPF is rectangular, and its impulse response is a Sinc function presented

11

Figure 3.2 Transmit spectrum mask for 20MHz transmission [7].

in the Fig. 3.3, whose support region extents to all past and future time. As a result, the

ideal LPF is impossible to realize, because it requires the knowledge of signals in the inﬁnite

future and past.

The practice of designing LPF is to approximate the ideal LPF by truncating and

windowing the inﬁnite impulse response, which results in a transition band that exhibits

a roll-oﬀ between the passband and the stopband. A ﬁlter is characterized by its cutoﬀ

frequency and rate of frequency roll-oﬀ which is expressed in dB/decade or dB/8ve. Signal

attenuation after cutoﬀ frequency depends on the design of the ﬁlter, speciﬁcally the order

of the ﬁlter. Fig. 3.4 illustrates the trade-oﬀ in designing LPF. As shown in the ﬁgure,

high-order ﬁlters have steeper roll-oﬀs, resulting in less distortion of wanted signal in the

passband, which is however at the cost of higher implementation complexity.

12

Figure 3.3 Impulse response of Rect function, Sinc(x).

Figure 3.4 Eﬀects of ﬁlter order on roll-oﬀ: Higher order ﬁlters have steeper transitions
and smaller transition bands.

13

−6−4−20246−0.4−0.20.00.20.40.60.81.0PassbandfsStopbandFrequency0.00.20.40.60.81.0Gainorder = 3order = 6order = 9Chapter 4

The Impacts of LPF Defects Under

the Microscope

In this chapter, we present microscopic measurements and models to characterize the

surprisingly notable impacts of LPF design on the performance of today’s 802.11 systems.

Using a bottom-up approach, we demonstrate how signal-level defects induced by LPF roll-

oﬀ propagate to the upper layers of wireless communication, reshaping bit error patterns

and degrading link performance.

The contents of this chapter were adapted from a published conference paper 1.

4.1 Signal-Level Characterization

In this section, we characterize the signal-level defects caused by LPF design. Specif-

ically, a basic trade-oﬀ in LPF design is to balance the steepness of frequency roll-oﬀ with

implementation complexity. As discussed before, a slow roll-oﬀ may distort wanted signal

components in pass-band, while a steep roll-oﬀ requires higher implementation cost. We

demonstrate the eﬀect of roll-oﬀ induced signal distortion in diﬀerent oﬀ-the-shelf chips,

and evaluate its signiﬁcance in practice in the presence of multi-path fading and mobile

1“Harnessing hardware defects for improving wireless link performance: Measurements and

applications”. Alireza Ameli Renani, Jun Huang, Guoliang Xing, Abdol-Hossein Esfahanian.

INFOCOM 2017 - IEEE Conference on Computer Communications.

14

environments.

4.1.1 Methodology

We characterize LPF design by measuring the frequency roll-oﬀ in the output signal

of oﬀ-the-shelf 802.11 chips. Existing measurement tools have several limitations for this

purpose. General-purpose spectrum analyzers can measure the power spectral density of

received signal, but cannot diﬀerentiate between signals of diﬀerent transmitters. Therefore,

in uncontrolled setups, their measurements can be polluted when the target signal mixes

with noise or interference. The ﬁrmwares of a few 802.11 chips can provide channel state

information (CSI) that characterizes signal-level distortions [33]. However, the reported CSI

is subject to the manipulation of CSI calibration algorithms, thus cannot accurately measure

frequency roll-oﬀ. The GNURadio/USRP platform [6] can be programmed to analyze raw

802.11 signals, but its results can be distorted by various processings along the receiving

pipeline of the USRP hardware.

To address these limitations, we carefully examine signal distortions introduced by

various analog and digital components of USRP, and then insert a compensation ﬁlter to

correct distortions before feeding the signal to the 802.11-compliant receiver. Fig. 4.1 shows

the architecture of the receiver pipeline used in our measurements, which uses a daughter-

bord of SBX. By checking the datasheet of each analog component of SBX, we ﬁnd that it

introduces ignorable distortion to the 802.11 signal on a 20 MHz channel. For example, the

LPF of SBX operates on a passband of 40 MHz. Thus the eﬀect of its roll-oﬀ on the 20 MHz

channel of 802.11 is non-signiﬁcant.

We ﬁnd that the major source of distortion inside the receiver is caused by the digital

down converter (DDC) of USRP, where the output of the ADC (100 million samples per

15

Figure 4.1 Architecture of the receiver front-end, consisting of a USRP2 and an SBX daugh-
terboard.

second (MS/s)) is down-sampled by a factor of 5 in order to generate a 20 MHz signal. To

suppress the aliasing caused by down-sampling, the DDC incorporates a 4-stage cascaded

integrator-comb (CIC) ﬁlter. Its magnitude response is given by

|Hpfq| “ | sinpπfq
sinp πf
R q

|4 « |R sinpπfq

πf

|4

(4.1)

where R is the rate change factor. The magnitude response is poor when R is odd, resulting

in signiﬁcant distortion to the received 802.11 signal. Fig. 4.2 illustrates the eﬀects of the

CIC ﬁlter on the received signals. As shown in the ﬁgure, the CIC ﬁlter causes a power

degradation as high as 18 dB at the border of the 802.11 20 MHz channel, and results in

a power loss of approximately 12 dB at the right-most and left-most data subcarriers. To

correct this distortion, the magnitude response of the compensation ﬁlter must be the inverse

of Eq. (4.1), which can be computed as

|Gpfq| «

ˇˇˇˇ

πf

R sinpπfq

ˇˇˇˇ4

16

(4.2)

Figure 4.2 Magnitude response of CIC ﬁlter for the USRP2.

We then compute the ﬁlter coeﬃcients by applying an inverse Fourier transform to Eq. (4.2),

and apply the ﬁlter on the signal received from USRP to compensate signal loss.

After compensating for the distortion caused by CIC ﬁlter, we measure frequency roll-

oﬀ using the long preamble of each received packet. Speciﬁcally, each long preamble of an

802.11 consists of known symbols of equal power transmitted on all subcarriers. As a result of

frequency roll-oﬀ, symbols transmitted on subcarriers close to the channel border may have

lower power. To quantify the power loss caused by roll-oﬀ, we ﬁrst correct the frequency

oﬀset between the transmitter and the USRP, and then normalize power measurements for

each subcarrier by computing the ratio with respect to the maximum power (measured in

dB), which can be expressed as

pi “ 10log10

1
n

ř

pi
n
k“1 pk

(4.3)

where n is the number of subcarriers and pi is the power of symbol transmitted on subcarrier

17

i.

4.1.2 Measuring Frequency Roll-Oﬀ

To obtain precise measurements of the frequency roll-oﬀ, we ﬁrst conduct experiments

in a controlled setup where transmitter and receiver are connected using a cable to exclude

the interference and distortion induced by wireless channel. Speciﬁcally, the 802.11 chip

is associated with an oﬀ-the-shelf access point (AP). The output signal of the 802.11 chip

is split into two ways using an RCA signal splitter. One way of the signal is fed to an

external antenna, which allows the 802.11 chip to communicate with the AP. The other way

is connected with an USRP2 via cable, such that the eﬀect of channel fading is excluded in

our measurement, allowing us to focus on the eﬀect of hardware-induced distortions. A step

attenuator is used to reduce the transmitter power by 10 dB, thus minimizing the risk of

any damage to the USRP2. We have checked the datasheets of the signal splitter and the

step attenuator to make sure they introduce ignorable distortions to signal. Fig. 4.3 shows

the setup of our measurement.

We measure the frequency roll-oﬀ induced by LPF in diﬀerent laptops and access

points. These devices use 802.11n chips manufactured by Broadcom, Qualcomm and Intel,

respectively (three 802.11 chip vendors that share a majority of the global market). The

BCM43xx series is widely used in mobile devices, including Apple MacBooks and iPhones,

while AR92xx chips are widely used in 802.11n access points of major vendors such as D-

Link, Linksys, TP-Link, and Netgear. The Intel 53xx series are popular in laptops, and

are widely used in academic research thanks to its capability of reporting channel state

information [33] [38] [46].

Fig. 4.4 shows the raw and compensated transmitter power spectral densities (PSDs)

18

Figure 4.3 Setup of over-cable measurements using an USRP and a 2ˆ2 MIMO AR9285
chip. The AR9285 is connected to an Agilent 8494B attenuator through an RCA DH24SP
2-way signal splitter, and the signal was measured from the main antenna port.

measured using ﬁve 802.11 chips, including Atheros’ AR2425 and AR9285, Broadcom’s

BCM4312 and BCM4331, and Intel5300. Roll-oﬀ in power spectral density caused by LPF

design is observed in all of the chipsets we studied. We also observe that signal power loss

at the channel border is diﬀerent across diﬀerent chips. For example, the AR2425 has a loss

of 2 dB on subcarrier -26, while BCM4312 shows a 4 dB loss on the same subcarrier. Such

diﬀerence is probably caused by the diﬀerent design of LPF across chips. As we expected, the

eﬀect of roll-oﬀ can be clearly observed on symbols received at border subcarriers. Moreover,

the CIC ﬁlter of USRP’s DDC introduces signiﬁcant additional distortions, which demon-

strates the necessity of careful compensation. In particular, the power degradation caused

19

Figure 4.4 Transmitter power spectral densities of diﬀerent 802.11 chipsets measured before
and after compensation for CIC roll-oﬀ.

by the roll-oﬀ of the transmitter’s LPF can be clearly observed even after compensating the

distortion of CIC ﬁlter. As an example, Fig. 4.5 compares the power loss on subcarriers -26

and +26 (border subcarriers) with that on subcarriers -1 and +1 (center subcarriers). As

shown in the ﬁgure, border subcarriers suﬀer up to 5 dB additional power loss because of

the roll-oﬀ of LPF.

The measurements shown in Fig. 4.4 and Fig. 4.5 only account for the loss caused

by the transmitter’s LPF. In practice, LPF in the receiver may introduce additional degra-

dations, which further distorts the PSD of received signal. Unfortunately, measuring the

distortions caused by the receiver’s LPF is not feasible, so we assume that the receiver’s

LPF roll-oﬀ is the same as that of the transmitter. Fig. 4.6 depicts the projected power

loss on subcarrier 26 for diﬀerent transmitter and receiver pairs. The power loss of border

subcarriers can be as high as 10 dB.

4.1.3 Signiﬁcance in Practical Environments

We have shown that the distortion caused by LPF roll-oﬀ is signiﬁcant in a controlled

setup where the eﬀect of channel fading is excluded. In the following, we study the eﬀect

20

Figure 4.5 Power losses caused by transmitter LPF on border subcarriers (subcarriers -26
and +26) and center subcarriers (subcarriers -1 and +1). Error bars show the minimum and
maximum measurement results.

of LPF roll-oﬀ in practical environments in the presence of multipath fading and device

mobility which causes frequency-selective fading and fast channel variation, respectively.

Our experimental setup consists of a client associated with an AP, which use AR2425 and

BCM4331, respectively. Another netbook equipped with Intel5300 is employed to measure

the PSD of signals.

To study the eﬀects of LPF roll-oﬀ in the presence of multipath fading, we deploy the

client in a position outside of the AP’s line-of-sight, with a distance about 15 feet between

them. The client and the AP are kept stationary during measurement. Fig. 4.7 shows

the PSD of the AP’s signal measured using the netbook equipped with Intel5300 at three

randomly selected positions inside an oﬃce building. As shown in the ﬁgure, the PSDs of

received signals diﬀer signiﬁcantly when measuring at diﬀerent positions. The deepest fading

21

Figure 4.6 Projected power losses on subcarrier 26 for diﬀerent transmitter and receiver
pairs. The results were computed based on the measurement shown in Fig. 4.5.

ranges from 12 dB to 30 dB, which are observed at diﬀerent subcarriers. In comparison,

power loss at border subcarriers becomes less signiﬁcant. Fig. 4.8 illustrates the CDFs of

symbol power measured on channel border (i.e., subcarrier -26 and +26) and channel center

(i.e., subcarrier -1 and +1) during the same experiment. In particular, we observe that the

power losses are similar on border and center subcarriers. The results indicate that in most

cases, the eﬀect of LPF roll-oﬀ is overwhelmed by the distortion caused by multipath fading.

To study the eﬀects of LPF roll-oﬀ in the presence of device mobility, we conduct

experiments using the same devices, and have the client move at a walking speed of about 3

mph. Fig. 4.9 shows the average PSD of the AP’s signal when the number of measurement

packets increases. The ﬁgure shows that, upon receiving 500 packets, the average power

losses on border subcarriers increase to 3 dB and 3.5 dB, respectively, while other subcarriers

experience lower power losses. The result indicates that, the LPF-induced roll-oﬀ becomes

signiﬁcant in the average PSD as the number of received packets increases. This is because

of the fast variation of wireless channel, which cancels the eﬀects of multipath fading in the

average of PSD after a short period of device movement. Fig. 4.10 shows how the average

22

 0 2 4 6 8 10 12AR242x/AR242xAR242x/AR9285AR242x/BCM4312AR242x/BCM4331AR242x/Intel5300AR9285/AR9285AR9285/BCM4312AR9285/BCM4331AR9285/Intel5300BCM4312/BCM4312BCM4312/BCM4331BCM4312/Intel5300BCM4331/BCM4331BCM4331/Intel5300Intel5300/Intel5300Power Loss (dB)Figure 4.7 Transmitter spectral densities measured at three randomly selected static links
within a single oﬃce building. The LPF-induced roll-oﬀ became insigniﬁcant as a result of
the severe distortion caused by multipath fading.

symbol powers converge with time. At the beginning of the measurement, both border

and center subcarriers exhibit similar power losses. After receiving 1500 packets, the eﬀect

of LPF roll-oﬀ can be clearly observed, as symbols received on border subcarriers suﬀer a

higher power loss than that received on center subcarriers. After 3000 packets, the average

power loss on center and border subcarriers converge to lower than 0.5 dB and above 2 dB,

respectively.

4.1.4 Summary

In summary, our signal-level measurements show that:

• In oﬀ-the-shelf 802.11 devices, the design trade-oﬀ of LPF causes notable signal-level

distortions, where subcarriers close to channel border suﬀer higher power loss due to

LPF roll-oﬀ.

• In the presence of multipath propagation, the eﬀects of frequency-selective fading over-

whelms the impacts of LPF roll-oﬀ on static links.

• In mobile environments, the eﬀects of frequency-selective fading can be averaged out

after a short-period of movement, making LPF roll-oﬀ the dominant cause of signal

23

-30-25-20-15-10-5 0-10-50+5+10Normalized Power (dB)        -10-50+5+10 Frequency (MHz)       -10-50+5+10  Figure 4.8 CDFs of power losses measured in the presence of multipath fading. The power
losses at channel center (i.e., subcarrier +1 and -1) and channel border (i.e., subcarrier +26
and -26) are similar, as the eﬀect of LPF roll-oﬀ is overwhelmed by multipath fading.

distortion.

4.2

Impacts on Upper Layers

In the following, we study how signal-level distortion caused by LPF roll-oﬀ propa-

gates to the bit-level and link layer, reshaping bit error patterns and degrading link reliability.

4.2.1 From Signal-Level to Bit-Level

Modern 802.11 adapters adopt OFDM in PHY, which transmits data on diﬀerent

subcarriers. To understand how LPF induced signal-level defects propagate to the upper-

layers of wireless communication, Fig. 4.11 gives a simple example and demonstrates how

LPF roll-oﬀ aﬀects the data transmitted on diﬀerent subcarriers. Speciﬁcally, when the eﬀect

of LPF roll-oﬀ is signiﬁcant, subcarriers at channel border will experience higher power losses,

24

−40−35−30−25−20−15−10−50Power (db)0.00.20.40.60.81.0CDFSubcarrier 1Subcarrier -1Subcarrier 26Subcarrier -26Figure 4.9 The average PSD measured on an 802.11 receiver moving at walking speed.
LPF-induced roll-oﬀ becomes increasingly pronounced as the number of received packets
increases. Roll-oﬀ is the most signiﬁcant error source in the average PSD measured using
500 packets.

leading to lower SNRs and higher error rates, which will result in a periodic error pattern

that repeats itself every OFDM symbol.

We note that although other components in the transceiver pipeline of 802.11 may also

aﬀect communication performance, the impact of roll-oﬀ induced signal distortion remains

signiﬁcant at bit-level. Speciﬁcally, in 802.11, data bits are ﬁrst scrambled, channel coded,

and interleaved before transmitting on diﬀerent subcarriers using OFDM. As we discussed

earlier, LPF roll-oﬀ will introduce a periodic error pattern to received coded bits. After

demodulation, coded bits are ﬁrst de-interleaved, where positions of coded bits are changed

pseudo-randomly. According to 802.11 standard, the interleaving depth equals OFDM sym-

bol size, and the interleaving map remains the same across all OFDM symbols. Therefore,

the eﬀect of interleaving is reshaping rather than eliminating the periodic error pattern of

coded bits. During channel decoding, whether a data bit can be correctly decoded or not

depends on the errors of coded bits. Therefore, errors in the output of channel decoder is

25

-30-25-20-15-10-5 0     Normalized Power (dBr) (a) 1 Packet-30-25-20-15-10-5 0       (b) 50 Packet-30-25-20-15-10-5 0       (c) 100 Packet-30-25-20-15-10-5 0-10-50+5+10Normalized Power (dBr)Frequency (MHz)(d) 200 Packet-30-25-20-15-10-5 0-10-50+5+10 Frequency (MHz)(e) 500 Packet-4-3.5-3-2.5-2-1.5-1-0.5 0-10-50+5+10 Frequency (MHz)Zoom of figure (e)Figure 4.10 The average of power losses measured at channel border (i.e., subcarrier +26
and -26) and channel center (i.e., subcarrier +1 and -1) as the number of received packets
increases.

determined by the error pattern of coded bits. Finally, the de-scrambler XORs decoded bits

with a pre-deﬁned pseudo-random sequence, which does not aﬀect bit errors. As a result, the

periodic error pattern caused by LPF roll-oﬀ will be transformed and observed at bit-level.

In the following, we ﬁrst conduct measurements to conﬁrm the existence of periodic

bit error pattern, and then examine its relation with LPF design based on simulations with

a standard-compliant software transceiving pipeline.

4.2.2 Measuring Bit Error Patterns

In this section, we conduct experiments to examine the bit error rate to study the

eﬀect of LPF roll-oﬀ in bit-level using Commercial Oﬀ-The-Shelf (COTS) 802.11 devices.

We modiﬁed an ath9k which is a free and open source software (FOSS) wireless driver for

Atheros IEEE 802.11n chipsets to report packets with CRC errors. That makes AR9285

26

050010001500200025003000Number of Packets−18−16−14−12−10−8−6−4−20Average Power (dB)Subcarrier 1Subcarrier -1Subcarrier 26Subcarrier -26Figure 4.11 Bits close to the border suﬀer from in-band signal loss.

report corrupted packets which allows us to use the device as receiver for measuring bit error

patterns. Other devices are employed as transmitters. We set up a Virtual Access Point

(VAP) on the receiver so we can control the transmission power and transmission rate. We

examine the eﬀect of the following factors in the study of bit-level patterns:

• Transmitter Chipset. Our experiment setup consists of four devices equipped with

AR9285, AR2425, AR9001, and Intel5100.

• Transmission Power. We conduct experiments using a wide range of transmission

powers, from minimum to maximum allowed on a chipset.

• Transmission Rate. We repeated the experiments using all the 802.11g rates.

• Distance. We conduct the experiments over a wide range of transmitter to receiver

distances.

• Payload. We use diﬀerent payloads such as loading all the bytes with 0xﬀ, 0x00, and

also with random values.

27

Figure 4.12 Byte error rates for transmission rate of 54Mbps. The vertical axis represents
the normalized bit error rate and the horizontal axis represents the byte position within the
payload.

Figure 4.13 Byte error rates for transmission rate of 48Mbps. The vertical axis represents
the normalized bit error rate and the horizontal axis represents the byte position within the
payload.

• Mobility. We conduct the experiments using stationary and mobile transmitter and

receiver.

We observe a periodic bit error pattern when the receiver receives enough errors

in conducting experiments using all the above factors. Fig. 4.12 and Fig. 4.13 plot byte

error rates of 100,000 packets transmitted by AR2425 device over a stationary link with

transmission rates of 54Mbps and 48Mbps respectively. The error rates of 8 bits are averaged

to a byte error rate to decrease the number of data points for plotting the entire payload.

We observe that bits at certain positions experience signiﬁcantly higher error rates. Similar

bit error patterns are also observed on other devices in all conﬁgurations.

In most of the experiments, the rate of the bit errors also increased in proportion to

the position of the bits in the payload. Han et al. referred to this pattern as slope pattern

[34][35]. They concluded that clock drift and changes in channel condition are the main

28

02004006008001000120014001600Byte Position0.0000.0010.0020.0030.0040.0050.0060.0070.008Byte Error Rate02004006008001000120014001600Byte Position0.0000.0010.0020.0030.0040.0050.0060.007Byte Error RateFigure 4.14 The autocorrelation of error rates over bit positions under diﬀerent transmis-
sion rates. The period of the error pattern equals to the number of bits carried by one
OFDM symbol. The transmission rates are 48Mbps and 54Mbps for left and right ﬁgures,
respectively.

reasons of the slope pattern.

We further validate the periodicity of the bit error pattern by computing the au-

tocorrelation of error rates over bit positions. Fig. 4.14 shows the results measured using

transmission rates of 48Mbps and 54Mbps. The peaks are identiﬁed at bit positions 192 and

216 respectively, which are equal to the number of bits carried by one OFDM symbol at

48Mbps and54 Mbps transmission rates. Table 4.1 shows the modulation, coding rate and

number of data bits per symbol for all the transmission rates in 802.11g. The results are

consistent with previous measurement study [34] [35], and are accordance with our analysis

discussed in Section 4.2.1, conﬁrming that the cause of the periodic bit error pattern is not

eﬀected by environmental factors and it’s rooted in the transceiver hardware.

29

Transmission
Rate(mbps)

Modulation Coded

bits

per carrier

6
9
12
18
24
36
48
54

1
BPSK
1
BPSK
2
QPSK
QPSK
2
16-QAM 4
16-QAM 4
64-QAM 6
64-QAM 6

Coded Bits
per OFDM
Symbol
48
48
96
96
192
192
288
288

Coding Rate Data Bits per

Symbol

1/2
3/4
1/2
3/4
1/2
3/4
2/3
3/4

24
36
48
72
96
144
192
216

Table 4.1 802.11g OFDM modulations, coding and data rates [31] [61].

4.2.3 Simulation-Based Diagnosis

We conduct simulation-based diagnosis using a pair of 802.11-compliant software

transceivers. Both transmitter and receiver are interfaced with a digital LPF, and then

connected via an AWGN channel of tunable SNR. To uncover the impacts of LPF roll-oﬀ,

we measure bit error patterns and packet error rates using extensive simulations and compare

the results with that measured using COTS 802.11 devices in testbed experiments. We then

tune the order and cut-oﬀ frequency of LPFs to diagnose their impacts at bit-level and link

layer.

Fig. 4.17 shows the bit error patterns measured in simulations and experiments at

54 Mbps. Similar results are observed under other settings, which are skipped here due to

space limit. The experiment result shown in Fig. 4.17a is obtained on a WiFi link consisting

of two AR9285 chips. Fig. 4.17b shows simulation result obtained under a SNR of 24 dB,

using a 5-order LPF with a cut-oﬀ frequency of 5 subcarriers at each edge of the channel.

The periodic bit error pattern can be clearly observed and is accordance with experiment

result. In comparison, Fig. 4.17c shows simulation results obtained under the same SNR,

30

but using a 31-order LPF with a cut-oﬀ frequency of only 2 subcarriers. As shown in the

ﬁgure, the periodic pattern is completely eliminated. This is because a LPF of higher order

and lower cut-oﬀ frequency will result in a sharp transition between passband and stopband,

and eﬀectively reduce the power loss of wanted signal caused by roll-oﬀ.

Figs. 4.15 and 4.16 show the eﬀects of cutoﬀ frequency and ﬁlter order on the packet

success rate when the SNR is 20dB. As shown in the ﬁgures, both cut-oﬀ frequency and

ﬁlter order have signiﬁcant impacts on packet loss rate. Low ﬁlter order makes the transition

between passband and stopband slower, which results in more signal distortions and packet

loss. In particular, changing the cutoﬀ frequency from 5 subcarriers down to 2 subcarriers

and increasing the ﬁlter order from 4 to 10 increases the packet success rate from less than

20% to more than 70%.

The simulation results also imply that LPF roll-oﬀ is the root cause of the periodic

bit error pattern observed on COTS 802.11 devices. The result demonstrate the signiﬁcant

impact of LPF design on link reliability.

31

Figure 4.15 Filter simulation results illustrating the eﬀect of cut-oﬀ frequency and ﬁlter
order on packet success rate.

Figure 4.16 Heat map of ﬁlter simulation results illustrating the eﬀect of cut-oﬀ frequency
and ﬁlter order on packet success rate.

32

Figure 4.17 Bit error rates measured using a) WiFi link of two AR9285 chips; b) simulation
of a 6-tap ﬁlter with cut-oﬀ frequency of 5 subcarriers; c) simulation of a 32-tap ﬁlter with
cut-oﬀ frequency of 2 subcarriers.

33

Chapter 5

Application in Video Streaming

The periodic bit error rate pattern has a high potential to get utilized in applications

which transmit data with diﬀerent levels of priority. Video streaming and forward error

correction (FEC) are examples of such applications. In this chapter, we evaluate the eﬀect

of utilizing the BER pattern on video streaming performance.

The contents of this chapter were adapted from a published conference paper 1.

5.1 The Unequal Error Protection Algorithm

According to the Cisco Visual Networking Index [1], video traﬃc will make up 80% of

all consumer Internet traﬃc by 2019. Mobile video traﬃc accounted for 55% of total mobile

data traﬃc in 2015. This traﬃc will increase by 11-fold by 2020, accounting for 75% of total

mobile data traﬃc. Unfortunately, the ﬂuctuations caused by the inherent unpredictability

of the wireless medium being used for streaming also causes ﬂuctuations in both the bit error

rates and throughput, resulting in choppy video playback [19].

A large body of work on video coding and streaming has explored various means

to improve the throughput, quality and performance of video streaming [39] [55] [19] [44].

Most of them are designed based on channel measurements and estimations which introduce

1“Harnessing hardware defects for improving wireless link performance: Measurements and
applications”. Alireza Ameli Renani, Jun Huang, Guoliang Xing, Abdol-Hossein Esfahanian. INFOCOM
2017 - IEEE Conference on Computer Communications.

34

processing overhead to the system.

In this section we present our state of the art unequal error protection algorithm

which increases the video streaming performance with minimal processing overhead and no

data overhead. Based on our study the processing overhead of our UEP algorithm is less

than the buﬀering time in the receiver, so it is not noticeable.

5.1.1 Background

Lightweight User Datagram Protocol or UDP-Lite is a protocol that delivers damaged

payloads to upper layers rather than discarding them. UDP-Lite does not drop corrupted

packets if the header does not contain errors. One of the major applications of UDP-Lite

is video streaming for the same reason. Receiving corrupted packets in video streaming are

useful and displaying them to the user is better than discarding them.

The H.264 codec [53] is the most widely used codec for video streaming [13]. We used

H.264 implemented in x264 [14]. The x264 has been used in many video services on the web

such as Youtube, Facebook, Vimeo and Hulu [14] [15]. H.264 includes three major type of

frames: I-frame, P-frame and B-frame. One I-frame and multiple P- and B-frames form a

single Group Of Pictures (GOP). The I-frame is the intra-coded picture and is the key frame

as it contains static image from a given scene. The P-frame is the predicted picture and only

holds changes made from one frame to another (i.e., only moving parts are encoded onto

P-frames). The B-frame is the bi-predictive picture and retains the diﬀerences between the

current frame and the two frames that precede and follow it.

Because the I-frame is required to decode all consecutive P- and B-frames within a

given GOP, it is necessarily the most valuable frame, since losing it would mean the loss of the

entire GOP. P-frames are second in line of priority, since B-frames need them for decoding.

35

This means B-frames have the lowest priority, as no other frame types are dependent on

them. Based on this observation, we assign diﬀerent levels of priority to video frame types.

I-frames have the highest priority among the frames, then P-frames and ﬁnally B-frames.

5.1.2 Basic Idea

Our basic idea to improve the video streaming quality is to transfer the data with

highest priority in payload byte positions with lowest error rate. Therefore I-frame data

should be positioned into payload bytes that experience the lowest error rates. From the

remaining byte positions, P-frames should take the bytes that experience lowest error rates.

Finally B-frames should take the remaining positions in the payload. Such reordering moves

the errors to less sensitive parts of the data which results in higher quality in the ﬁnal result.

Based on this idea, we plan to develop our UEP algorithm to increase the quality of the

video streaming without increasing data overhead.

We explain the display order of the frames in MPEG compression by an example.

Assuming GOP = 12 with two B frames between two P frames or between an I frame and

a P frame, the display sequence of a GOP will be:

I, B1, B2, P1, B3, B4, P2, B5, B6, P3, B7, B8

(5.1)

The order of the frames change in standard video streaming and is diﬀerent from

their display order because of the dependencies between them. The idea is to stream the

frames that are dependent to other frames last as they should get decode last. So the order

36

of the frames in the standard data stream will be:

I, P1, B1, B2, P2, B3, B4, P3, B5, B6, I2, B7, B8

(5.2)

In the sequence above, I2 is the I-frame for the next GOP which comes before B7 and B8

because both B-frames are dependent to that I frame.

Our algorithm reorders the above data stream based on priority of the frames.

I

frames have the highest priority because all P and B frames need them to decode. Next

are the P frames as they only need the I frames to decode. Finally B frames are the least

important in the GOP as no other frames depend on them. Here is the data stream after

reordering the frames:

I, P1, P2, P3, B1, B2, B3, B4, B5, B6, B7, B8

(5.3)

The above step can be reversed in the receiver to reorder the data back to the original

order so they can get decoded and displayed to the user. The order of the frames becomes 5.2

after reverting the changes made by the transmitter. The rest of the process will be the same

as standard streaming.

5.1.3 Algorithm Design

The goal of our unequal error protection algorithm is to increase the quality of the

video streaming without adding extra data overhead and minimizing the processing overhead.

Given the BER pattern as the input of the algorithm, the output is the mapping of the video

bytes into wireless frames.

37

We make two contributions in the design of UEP algorithm:

• Reordering the payload based on byte error rate. UEP algorithm takes all the required

wireless packets into account and sorts the payload bytes based on their error rate

which is known to both transmitter and receiver ahead of time, and moves the highest

priority data bytes into positions that experience the lowest error.

• Video frame based FEC. In wireless communication, errors usually happen in bursts

so instead of using FEC to wireless packets, our UEP algorithm uses FEC on video

frames before reordering. This causes the burst of errors in a wireless packet to get

distributed among diﬀerent video frames which have diﬀerent FEC data hence results

in more error correction capability.

The UEP algorithm starts by sorting the payload bytes by their byte error rates which

is known prior to streaming by both transmitter and receiver. Next it reads and decodes all

the frames in a GOP, then it sorts them based on their priority. Each GOP will most likely

get transmitted using more than one wireless packet, so the size ratio of I frame in GOP

should be calculated in order to map the I frame data in that ratio of each packet payload.

The same goes for P and B frames. Next, the UEP algorithm maps the highest priority

bytes (I-frame bytes ﬁrst) to the payload bytes with lowest error rates across all the wireless

packets.

Fig. 5.1 presents the original video frame positions in wireless packets. In this exam-

ple, the GOP consists of one I frame, one P frame and two B frames and the GOP requires

ﬁve wireless packets to transmit all the video frames. Fig. 5.2 presents the how UEP algo-

rithm orders the same GOP using the same number of wireless packets based on their priority

to protect them against periodic errors. Fig. 5.3 illustrates the mapping of higher priority

38

Figure 5.1 Standard video streaming is used to transmit a GOP with one I-frame, one
P-frame and two B-frames using ﬁve wireless packets.

Figure 5.2 UEP algorithm is used to transmit a GOP with one I-frame, one P-frame and
two B-frames using ﬁve wireless packets.

data to payload positions with lower error rates in just one wireless packet. The receiver

goes through the steps in reverse and outputs the original video frame order illustrated in

Equ.5.2. The pseudocode of the UEP algorithm is described in Algorithm 1.

The UEP algorithm distributes errors in a single wireless packet to multiple video

frames. To maximize the eﬀect of this scheme, we further incorporate a video-frame-based

FEC. In wireless communication, errors usually happen in bursts, so using FEC on video

frames before reordering causes errors to get distributed among diﬀerent video frames in

the same wireless frame which leads to more error correction. For the purpose of the video

streaming application, we decided to use a popular Reed-Solomon code RS(255,223) for both

39

I frameI frameI frameP1frameB1frameB2frameWireless FramesFrame Bytes12345Wireless Frames12345I framePframeB2frameB1frameFrame Bytes Sorted based on Error RateFigure 5.3 UEP algorithm maps higher priority data to bit positions that experience lowest
error rates in a wireless packet.

standard and UEP streamings. RS(255,223) adds 32 parity bytes for each 223 bytes of data

to a sum of 255 byte block and it can recover up to 16 corrupted bytes in each data block.

5.2

Implementation

In this section we present the details of the implementation of the UEP algorithm. We

use trace-driven simulation to evaluate the improvements resulting from our UEP algorithm

compared to standard video streaming. The experiment setup consists of two netbooks with

Atheros AR2425 and AR9285 chipsets. There is no need to modify the ﬁrmware in order

to report corrupted frames to the user space, compared to some chipsets like Broadcom

which require a modiﬁcation to an open source ﬁrmware like OpenFWWF [32], to pass

the corrupted frames to user space. Although some drivers require minimal modiﬁcation.

Simply removing the condition to drop frames containing CRC errors in the ath9k drivers is

suﬃcient to enable all Atheros chipsets with that driver to support our UEP algorithm. We

40

theorize that other drivers that drop frames containing CRC errors would also require similar

modiﬁcation to support our algorithm, but this would be proven in our future work. It should

be noted that there are other drivers currently available that support our UEP algorithm,

such as the driver used by in [34] [35]. For these experiments, we installed the modiﬁed ath9k

wireless driver mentioned in 4.2.2 to have the driver report all packets including corrupted

ones that would fail the CRC, to the user space. We must state that the simple modiﬁcation

we made in the driver does not aﬀect the generality of our evaluation results, and our UEP

algorithm can deliver the desired improvement as long as the error pattern exists. Although

the UDP-lite protocol was speciﬁcally designed for use in cases where receiving corrupted

frames was preferential to dropping them, it is still not fully supported, and lower layers still

discard packets containing CRC errors. However, if ﬁrmwares and drivers support UDP-lite

in the future, our UEP algorithm will then be viable for use on all chipsets.

We use ﬀmpeg multimedia framework [5] to encode, decode, transcode and play the

video ﬁles. The ﬀmpeg command line tool is used for transcoding and the ﬀmpeg libraries

are used to decode the input ﬁles. After executing our UEP algorithm and masking the data

with our traces to simulate streaming, we use the ﬀmpeg libraries to encode the resulting

video ﬁles into yuv formatted videos. We further process the outputs using the VQMT

tool [11], which is a program for objective video quality assessment.

5.3 Evaluation

The primary focus of this section is to compare traditional video streaming to stream-

ing using UEP algorithm. A major advantage of our UEP algorithm is that it can be com-

bined with any physical or MAC layer schemes.

41

5.3.1 Setup and Conﬁguration

We use trace-driven simulations to evaluate both traditional video streaming and

UEP streaming. The experiment setup to collect traces consists of an Eee PC with Atheros

AR9285 chipset as the receiver and an Acer aspire with Atheros AR2425 chipset as transmit-

ter. The transmitter has a virtual access point (VAP) setup on it, thus enables us to control

the transmission rate. As discussed in 4, the transmission rate aﬀects the period of the BER

pattern, making it crucial to keep it at a constant rate for our experiments. Using the same

trace for standard and UEP streamings makes the comparison fair in all aspects. We leave

studying the eﬀect of dynamic transmission rate on the video streaming performance for our

future work.

For trace collection, we use the ping command to generate wireless frames with max-

imum payload size of 1464 bytes. We set the protocol to udp-lite as video streaming and

multimedia communication in general is one of the major applications of that protocol. We

collect 10,000 packets in each trace. The payload is set to 0xﬀ for all the bytes of the payload.

According to [34][35] and our analysis in 4, the payload data does not have any eﬀects on

the bit error rate. The major advantage of using 0Xﬀ as payload is that any 0 bit in the

bit stream is an error bit, so a simple XNOR of the payload with video bit stream in the

next step aﬀects the appropriate bit in the video stream simulating the actual transmission.

We apply the same trace to both traditional and UEP video streams and then evaluate the

results.

In all experiments the receiver is mobile and traveling with walking speed within the

15 feet range of the transmitter in a 300 square feet room, representing a mobile device in

an oﬃce environment.

42

PSNR (dB) MOS

ą 37
31 - 37
25 - 31
20 - 25
ă 20

5 (Excellent)
4 (Good)
3 (Fair)
2 (Poor)
1 (Bad)

Table 5.1 PSNR to MOS conversion [41] [51].

We use the well-known Akiyo video ﬁle for these experiments. The ﬁle format is mp4

and the video is 10 seconds long. The frame rate is 30fps so the video ﬁle has 300 frames in

total. The GOP size of the ﬁle is 30.

5.3.2 Baseline and Metric

We utilize PSNR, the most widely used metric, to quantify video quality. PSNR is

purely mathematically deﬁned metric and some literature recommend using metrics that

have similar characteristics of Human Visual System (HVS) [42]. For that reason, we also

use the Structural SIMilarity (SSIM) index to compare the video quality. The last metric

we use is the mean opinion score (MOS) which is the perception of video quality to human

eyes. MOS expresses the video quality in ﬁve levels: excellent, good, fair, poor and bad.

Table 5.1 shows the PSNR to MOS conversion.

5.3.3 Results

Fig. 5.4 shows one video frame recorded using traditional video streaming with PSNR

of 12, SSIM of 0.57 and based on Table 5.1 falls into the Bad category in MOS. Fig. 5.5

shows the exact same frame recorded using our UEP streaming in the same experiment. The

measured PSNR is 30, SSIM is 0.93 and it falls into Fair category in MOS. The improvement

43

Figure 5.4 A video frame of akiyo.mp4 recorded from standard streaming. Many parts of
the frame are completely corrupted. With PSNR of 12 it falls into the lowest MOS category.

of the quality in UEP streaming is obvious to human eye, but all the metrics validate it as

well.

In standard streaming most of the video frame parts are either completely lost or

extremely corrupted, but the UEP algorithm, with minor corruptions, is able to deliver

impressively higher quality video over the same unreliable wireless channel.

Fig. 5.6 plots the PSNR for every frame in the akiyo.mp4 video streamed using both

traditional and UEP streamings. We repeat the experiment 10 times using 10 diﬀerent traces

and average the PSNRs for each frame. As shown in the ﬁgure, the PSNRs range from 20

dB to 70 dB in traditional streaming while UEP keeps the PSNR constantly at or above 70

dB.

Fig. 5.7 presents the histogram of frame qualities for the same experiment. As shown

in the ﬁgure, UEP streaming produces fewer frames with bad, poor, fair and good rat-

ings, and 40% more frames with excellent rating compared to traditional streaming. The

video streaming experiment results indicate that our UEP algorithm surpasses the standard

44

Figure 5.5 A video frame of akiyo.mp4 recorded from UEP streaming which shows major
improvements in quality of the video. With PSNR of 30, it is in the high range of fair
category in MOS.

streaming by a clear margin without adding extra data overhead.

The average SSIM index of all the video frames in the abovementioned experiment

for standard and UEP video streamings are 0.93 and 0.96, respectively. The UEP algorithm

shows massive improvements in video streaming according to all three metrics we use in this

work.

We discussed the aggregated results which show UEP algorithm is successfully moving

the bit errors to less sensitive parts of the video. We compared the results with standard

streaming by using the same trace for both methods and UEP algorithm achieved extremely

higher quality videos. Next we discuss the individual experiment results and also the eﬀect

of video frame based FEC and its impact on correcting the burst of errors distributed by the

UEP algorithm to diﬀerent video frames.

Table 5.2 shows the average PSNR and SSIM in all the individual experiments and

also the overall average of the 10 experiments conducted in this section.

45

Figure 5.6 PSNR of all 300 frames in akiyo.mp4 streamed using UEP and standard stream-
ings.

Figures 5.8 to 5.27 plot the PSNR and MOS quality levels for every frame in the

akiyo.mp4 video streamed using both traditional and UEP streamings in the rest of the

experiments.

Fig. 5.8 plots the PSNR for every frame in the akiyo.mp4 video streamed using

both traditional and UEP streamings in experiment number 1. The ﬁgure shows that UEP

algorithm is able to deliver all 300 video frames with PSNR of 100 and in contrast traditional

streaming has only 154 video frames with PSNR of 100. As UEP algorithm only relocates the

errors, we can conclude that it is the video frame based FEC that is able to correct the errors

that appeared in almost 50% of the video frames and it corrected all video frames to achieve

PSNR of 100. Fig. 5.9 presents the histogram of frame qualities for the same experiment.

The ﬁgure indicates that the UEP is successfully distributing the burst of errors even in poor

video frames and video frame based FEC is able to correct all those errors and deliver 100

46

Figure 5.7 MOS of all the frames in akiyo.mp4 streamed using UEP and standard stream-
ings.

PSNR for all the video frames. Figures 5.10 and 5.11 present the result of the experiment

number 2 which has a highly corrupted trace. Even in the extreme case scenario which most

of the video frames streamed using standard streaming fall into the bad MOS, the UEP is

able to deliver most of the video frames with fair MOS, which is two level improvement

compared to standard streaming.

Figures 5.12 to 5.15 present the result of the experiment numbers 3 and 4. These two

experiments are the only experiments that standard streaming is delivering slightly higher

quality video frames. The average PSNR of all the frames in experiment 3 is 24 and 21 for

standard and UEP streamings. The values are 23 and 21 for standard and UEP streamings

in experiment 4. This is expected as the goal of the UEP algorithm is to increase the average

performance and does not guarantee higher quality in every situation. In chapter 4 we prove

that the periodic bit error pattern exists in every wireless communication, no matter what

47

PSNR

SSIM

Experiment UEP
100
27.94
26.02
25.8
100
100
100
100
100
100

1
2
3
4
5
6
7
8
9
10

Average

77.976

1

0.89
0.87
0.87

1

STD UEP STD
0.97
64.91
20.22
0.76
0.86
25.6
26.72
0.86
81.02
61.88
48.4
79.15
65.41
32.79
50.61

0.99
0.99
0.99
0.99
0.95
0.936

1
1
1
1
1
1

0.963

Table 5.2 Average PSNR and SSIM of all the experiments.

environmental variables are. We expect anomalies and sometimes they dominate the periodic

bit error pattern, and that explains the weak performance of UEP algorithm in these two

experiments.

Figures 5.16 to 5.27 present the experiments 5-10. The UEP streaming is delivering

all the video frames with PSNR of 100 while the average PSNR of standard streaming reaches

below 32 in experiment 10. This shows a signiﬁcant improvement of video quality achieved

by combining the state of the art UEP algorithm to distribute the bit errors to diﬀerent

video frames, with the novel idea of applying FEC on video streaming which corrected all

the bit errors in all the video frames.

We list the FEC code we use in streaming in Chapter 5. It is a popular Reed-Solomon

code RS(255,223) for both standard and UEP streamings. RS(255,223) adds 32 parity bytes

for each 223 bytes of data to a sum of 255 byte block and it can recover up to 16 corrupted

bytes in each data block. We can use more parity bytes to increase the number of bytes

the system can correct. In the future work, we can study diﬀerent coding schemes and the

eﬀect of changing that threshold on system performance, and possibly use a dynamic FEC

48

Figure 5.8 PSNR of all 300 frames in akiyo.mp4 streamed using UEP and standard stream-
ings in experiment number 1.

to minimize the redundancy and maximize the error correction depending on the number of

errors.

49

Figure 5.9 MOS of all the frames in akiyo.mp4 streamed using UEP and standard stream-
ings in experiment number 1.

50

Algorithm 1 Pseudocode of the UEP algorithm.
1: max pkt size = 1500
2: error rates = sort(read error rates())// read and sort BERs
3: gop = read video frames()// read video frames
4: p, b = 0, 0
5: for i “ 0 Ñ lenpgopq do

6:
7:
8:
9:
10:
11:

if gop[i].type() == ”I” then

I frame = gop[i]

else if gop[i].type() == ”P” then

P frames[p++] = gop[i]

else if gop[i].type() == ”B” then

B frames[b++] = gop[i]

12: wireless pkts count = len(gop)/max pkt size
13:
14:
15: i = 0
16: reordered gop[i++] = I frame
17: for j “ 0 Ñ lenpP f ramesq do
19: for j “ 0 Ñ lenpB f ramesq do

reordered gop[i++] = P frames[j]

18:

20:

reordered gop[i++] = B frames[j]

21:
22:
23: data = get byte stream(reordered gop)
24: k = 0
25: for i “ 0 Ñ max pkt size do

for j “ 0 Ñ wireless pkts count do

26:
27:

29:

packets

uep data[error rates[i]][j] = data[k++]

30: for i “ 0 Ñ max pkt size do

31:

transmit(uep data[:][i])

51

Ź reorder gop based on frame priority

Ź get bytestream of the data

Ź transmit packet i

28: Ź put data with highest priority to byte position with lowest error rate across all the wireless

Figure 5.10 PSNR of all 300 frames in akiyo.mp4 streamed using UEP and standard stream-
ings in experiment number 2.

Figure 5.11 MOS of all the frames in akiyo.mp4 streamed using UEP and standard stream-
ings in experiment number 2.

52

Figure 5.12 PSNR of all 300 frames in akiyo.mp4 streamed using UEP and standard stream-
ings in experiment number 3.

Figure 5.13 MOS of all the frames in akiyo.mp4 streamed using UEP and standard stream-
ings in experiment number 3.

53

Figure 5.14 PSNR of all 300 frames in akiyo.mp4 streamed using UEP and standard stream-
ings in experiment number 4.

Figure 5.15 MOS of all the frames in akiyo.mp4 streamed using UEP and standard stream-
ings in experiment number 4.

54

Figure 5.16 PSNR of all 300 frames in akiyo.mp4 streamed using UEP and standard stream-
ings in experiment number 5.

Figure 5.17 MOS of all the frames in akiyo.mp4 streamed using UEP and standard stream-
ings in experiment number 5.

55

Figure 5.18 PSNR of all 300 frames in akiyo.mp4 streamed using UEP and standard stream-
ings in experiment number 6.

Figure 5.19 MOS of all the frames in akiyo.mp4 streamed using UEP and standard stream-
ings in experiment number 6.

56

Figure 5.20 PSNR of all 300 frames in akiyo.mp4 streamed using UEP and standard stream-
ings in experiment number 7.

Figure 5.21 MOS of all the frames in akiyo.mp4 streamed using UEP and standard stream-
ings in experiment number 7.

57

Figure 5.22 PSNR of all 300 frames in akiyo.mp4 streamed using UEP and standard stream-
ings in experiment number 8.

Figure 5.23 MOS of all the frames in akiyo.mp4 streamed using UEP and standard stream-
ings in experiment number 8.

58

Figure 5.24 PSNR of all 300 frames in akiyo.mp4 streamed using UEP and standard stream-
ings in experiment number 9.

Figure 5.25 MOS of all the frames in akiyo.mp4 streamed using UEP and standard stream-
ings in experiment number 9.

59

Figure 5.26 PSNR of all 300 frames in akiyo.mp4 streamed using UEP and standard stream-
ings in experiment number 10.

Figure 5.27 MOS of all the frames in akiyo.mp4 streamed using UEP and standard stream-
ings in experiment number 10.

60

Eﬀect of retransmission. Thus far, we showed the impressive eﬀect of UEP al-

gorithm in increasing the video streaming performance without considering retransmission.

In this section, we study the eﬀect of packet retransmission on the video quality in both

UEP and traditional streamings. Retransmission makes it possible for both streamings to

retransmit the wireless packet in case of failure. As the UEP algorithm has higher chance of

correcting the errors, this gives the standard streaming an advantage to close the quality gap

and deliver video frames with higher qualities, possibly very close to what UEP algorithm

can achieve. To study the eﬀect of retransmission in video quality, we conduct experiments

with 1, 2 or 3 retransmission possibilities.

Table 5.3 presents the average byte errors in I, P and B frames in all 10 experiments

without considering retransmission. It also presents the total errors observed in both stan-

dard and UEP streamings. The UEP streaming is experiencing approximately 50% fewer

errors than standard streaming which is signiﬁcantly lower considering all experiments are

conducted over identical traces.

Table 5.4 presents the average byte errors in I, P and B frames in all 10 experiments

by having 3 possible retransmissions for every wireless packet. Each algorithm retransmits

the wireless packet if the packet get delivered corrupted. The table shows much lower errors

in both standard and UEP streamings, and a closer gap in the total errors. UEP algorithm is

experiencing 20% fewer errors than standard streaming which is still signiﬁcant considering

UEP algorithm shifts errors from sensitive parts of the video to less sensitive parts. As an

example, the total errors in the I frames are 1700 and 502 in standard and UEP streamings,

respectively. This means UEP is experiencing 70% fewer errors in the most sensitive parts

of the video which shows signiﬁcant improvement in video streaming quality.

61

Frame type

STD

I frames
P frames
B frames

Total errors

14255.9
18102.8
17260.8
49619.5

UEP
5979.5
9346.1
9343.4
24669

Table 5.3 Average byte errors in I, P and B frames in standard and UEP streamings without
retransmission.

Frame type

I frames
P frames
B frames

Total errors

STD
1700.8
6269.3
4945.7
12915.8

UEP
502.5
3214.3
6505.9
10222.7

Table 5.4 Average byte errors in I, P and B frames in standard and UEP streamings with
3 retransmissions.

62

Chapter 6

Application in Vehicular

Communication

Recent years have witnessed the emergence of vehicular networks deployed for smarter

and safer transportation services, such as vehicle safety communication systems, context-

aware drive tuning, remote diagnosis and network-assisted autonomous driving. These ap-

plications typically impose stringent performance requirements on vehicular networks, in-

cluding low delivery delay, high packet reception ratio and spectrum eﬃciency. To support

performance-demanding applications in vehicular networks, previous research mainly focuses

on designing communication protocols and architectures that address the fast variation of

the wireless channel [21] [30] [48]. However, the eﬀects of hardware defects have been largely

unexplored, which leads to communication behaviors and performance issues at upper-layers

that are diﬃcult to diagnose and control [38] [34].

6.1 Background

With recent emerging growth in autonomous driving, Wireless Access in Vehicular

Environments (WAVE) is receiving a great interest. There are numerous beneﬁts for WAVE,

but we focus on safety aspect of it in Vehicular Ad hoc Networks (VANETs). WAVE is facing

many challenges mostly due to extremely fast varying channels hence it is crucial that packets

63

get delivered successfully specially in emergency scenarios.

There has been a great body of work in simulation and SDR based study of WAVE

which we discussed in Chapter 2. Recent developments of autonomous vehicles are bringing

more attention to Dedicated Short Range Communications (DSRC) in VANETs, opening

a huge market for wireless chipset vendors to make 802.11p chipsets. Recent studies have

focused on using 802.11a chipsets instead of 802.11p chipsets for two main reasons:

1. IEEE 802.11p standard [17] is based on IEEE 802.11a standard [16] with diﬀerences

in bandwidth, carrier spacing and symbol length. In 802.11p the channel bandwidth of

10MHz is usually used, with the optional 20MHz mode. The carrier spacing is reduced

by half compared to 802.11a. The symbol length is doubled.

2. There are no 802.11p chipsets in production at the time of writing this dissertation.

There are a few products available, but they are not mass produced and the vendors

do not sell them directly to costumers.

Almost every study that used 802.11a chipsets modiﬁed them to make them similar to

802.11p standards. However, these modiﬁed chipsets suﬀer from at least omitting one of the

speciﬁcations because of various hardware and software limitations, such as using 20MHz

channel instead of 10MHz channel, or using 2.4GHz band instead of 5.9GHz band of 802.11p

which ranges from 5.850 to 5.925 GHz.

There are a few prototypes for On Board Units (OBUs) and Road Side Units (RSUs)

from vendors such as Cohda Wireless [3], Unex [2] and DENSO [4] which make it possible

to conduct ﬁeld tests and execute measurement experiments.

64

6.2 Harnessing Hardware Defects in Vehicular Com-

munication

IEEE 802.11p standard is based on IEEE 802.11a, so our measurements discussed in

Chapter 4 is applied in 802.11p. Building on our measurements and models we conclude

that our UEP algorithm presented in Chapter 5 can be modiﬁed to harness hardware design

tradeoﬀs and improve DSRC throughput by introducing minimal processing overhead, po-

tentially saving lives in emergency scenarios. FEC can also be employed to reduce the errors

caused by fast varying channels in vehicular environments.

We design an architecture to exploit the systemic impacts of transceiver hardware

defects in vehicular networks. By modeling the systematic impacts of hardware defects, the

system can be employed for directing the tuning of various analog and digital components,

given the system performance requirements imposed by network applications. In addition,

it harnesses hardware defects to improve communication performance, leveraging the fact

that hardware induced error patterns are invariant in despite of the fast channel variation

in vehicular networks.

We describe the architecture of the system next, and then explain the three main

components of the system, hardware defect modeling, harnessing hardware defect at PHY

and exploiting BER pattern at link layer. Finally based on our proposed methodology and

architecture, we explain how one may implement the experiments and evaluate the results.

65

Figure 6.1 The architecture of the system.

6.3 Architecture

As shown in Fig. 6.1, our system consists of the following major components. The

hardware defect modeling models the error patterns caused by hardware defect and exposes

them to PHY and link layer components. At PHY, the system leverages the pattern to

improve channel equalization and decoding performance. At the link layer, our system

leverages error patterns for bit error estimation and uses the results to optimize various link

layer protocols, such as data rate adaptation and transmission power control.

6.3.1 Hardware Defect Modeling

Hardware defect modeling component models the error pattern based on communi-

cation theory rather than mathematical modeling. Recent research shows that accurate

per-packet BER estimation has the potential of greatly improving the performance of wire-

less networks [28]. As an example, in autonomous driving systems, sensory data streaming

techniques can tolerate a certain level of errors, where a corrupted packet can be recovered

66

using application layer code if the BER of the packet is lower than a threshold. Exposing

the error patterns to PHY and link layer protocols to enable ﬁne-grained performance opti-

mization. Conducting a systematic measurement study is required to gain insights into the

characteristics of vehicular network traﬃc in diﬀerent applications and scenarios in 802.11p

DSRC based V2V communications. Based on the results, eﬃcient algorithms can be devel-

oped to model the error pattern.

6.3.2 Harnessing Hardware Defects at PHY

If PHY knows the signal distortions prior to receiving the signal, it can optimize the

channel equalization and decoding algorithms to improve their performance. We tackle two

major challenges at the PHY of vehicular network stack. First, wireless communication in

vehicular networks suﬀers from noise and interference signiﬁcantly more than conventional

wireless networks. Second, communication in vehicular networks is subject to the fast fading

of wireless channel. Conventional receiver obtains channel estimation using packet preamble

at the beginning of packet receiving. However, such estimation may be quickly outdated in

vehicular environment due to the fast movements of vehicles, causing substantial bit errors.

The proposed new system addresses the above challenges using novel system designs that

integrate BER model from hardware defect modeling and signal distortion modeling.

6.3.3 Exploiting BER Pattern at Link Layer

The link layer can exploit the BER pattern models to signiﬁcantly improve the per-

formance of various link layer protocols such as bit rate adaptation, power control, selective

retransmission and hybrid automatic repeat request (ARQ). It can improve quality of service

67

while reducing spectrum overhead. Since this component operates at the link layer, it can

be directly implemented on oﬀ-the-shelf devices

6.4 Suggested Evaluation Methodology

Based on our proposed architecture, one may evaluate the system with a combined ef-

fort of theoretical analysis, simulation study and real testbed experiments. Publicly available

datasets [27] [18] can be used to evaluate the performance of hardware defect modeling. Real

channel traces can be fed into simulation tools for realistic and extensive simulation study

of the proposed algorithms and protocols. The performance of the system in real wireless

network testbeds can be also evaluated. The BER estimator and link layer protocols can be

implemented and evaluated on commercial oﬀ-the-shelf devices. A testbed of laptops and

smartphones may be required to evaluate the impact of the system on link layer throughput

and delay under mobility. One may conduct the following real testbed experiments to study

the systemic eﬀect of hardware defects in vehicular communication.

To obtain precise measurements of the frequency roll-oﬀ caused by the OBU, exper-

iments should be conducted in a controlled setup where transmitter and receiver are con-

nected using a cable to exclude the interference and distortion induced by wireless channel.

We presented our controlled experiments we conducted for 802.11g chipsets in Section 4.1.2.

The setup should consist of an 802.11p prototype and a USRP2. Fig. 6.2 shows en

example of such setup which consists of a Cohda MK5 OBU and a USRP2 and an XCVR2450

daughterboard. The Cohda MK5 is connected to the USRP2 through a Mini-Circuits VAT-

20+ SMA ﬁxed attenuator to reduce the transmission power by 20 dB, minimizing the risk

of any damage to the USRP2 or the daughterboard. We have checked the datasheet of such

68

Figure 6.2 Setup of over-cable measurements using an USRP2 and Cohda MK5 OBU. The
MK5 is connected to a Mini-Circuits VAT-20+ attenuator, and the signal was measured
from the antenna port 1.

attenuator to make sure it introduces ignorable distortions to signal.

After conducting the experiments in a controlled setup to measure the eﬀect of the

transmitter’s defects on the signal, one should conduct another set of controlled experiments

using two 802.11p prototypes such as Cohda MK5 OBU, to determine the eﬀect of the

receiver’s defects on the signal. The two prototypes should be connected using a cable to

avoid any distortions caused by the wireless channel. An attenuator such as Mini-Circuits

VAT-20+ should be used to reduce the risk of damage to the prototypes.

In this set of

experiments the signal passes through both transmitter and receiver, and it gets aﬀected by

69

Figure 6.3 Setup of measurements using two MK5 units over a stationary link. The antenna
is MGW-303 which includes two WiFi antennas and a GPS antenna.

the defects of both hardware, without any distortions caused by the wireless channel. By

comparing the results with the other controlled experiment, one can determine the eﬀect

of the receiver’s defects on the signal which is extremely hard to measure and requires

specialized measuring tools and techniques.

Next experiment should be conducted over a static wireless link to measure the eﬀect

of wireless channel on the bit error rate. To connect the two 802.11p prototypes over a

wireless link, one can use MGW-303 antenna which includes two WiFi antennas and a GPS

antenna. This experiment simulates a communication between two vehicles when they are

stationary. Fig. 6.3 shows an example of such setup.

70

The ﬁnal experiments should represent the extreme case scenario which is when two

vehicles are moving and communicating with each other. The same two devices as the last

experiment can be used to study the eﬀect of fast varying channels of vehicular networks on

bit error rates. It is preferred to use an antenna with a magnetic mount such as MGW-303

antenna. It makes it possible to mount it on the roof of the car to minimize the antenna

movement and vibration while driving, minimizing the noise caused by the movement. One

can also collect traces on these experiment to evaluate the UEP algorithm and study its

impact on improving the wireless communication in vehicular networks. We conducted the

the last two experiments on 802.11g chipsets in Section 4.1.3.

71

Chapter 7

Conclusion

We present an in-depth study of the surprising impacts of LPF defects on the per-

formance of today’s 802.11 systems. Based on microscopic measurements and models, we

show that LPF roll-oﬀ causes notable signal-level defects in oﬀ-the-shelf 802.11 chips, and

illustrate how the defects propagate to the upper-layers of wireless communication, reshaping

bit error patterns and degrading link performance.

Our results show that the impacts of LPF defects are particularly pronouncing on

mobile links, where LPF roll-oﬀ becomes the dominant source of error when the eﬀects of

varying channel are averaged out after a short period of movement. Driven by this insight,

we develop a novel unequal error protection algorithm that reorders video frame bytes based

on the error pattern caused by LPF defects, which substantially enhances video streaming

performance in mobile environments. A major advantage of our UEP algorithm is that it

can be combined with any physical or MAC layer schemes. It also has a great potential to

be used in forward error correction, channel coding, and many other applications to bring

about improvements on mobile links.

72

REFERENCES

73

REFERENCES

[1] Cisco visual networking index, http://www.cisco.com/.

[2] Cohda wireless, https://unex.com.tw/.

[3] Cohda wireless, http://www.cohdawireless.com/.

[4] Denso, https://www.denso.com/.

[5] ﬀmpeg, http://www.ﬀmpeg.org/.

[6] The gnu software radio, http://gnuradio.org/trac.

[7] IEEE Standard for Information technology-Telecommunications and information ex-
change between systems-Local and metropolitan area networks-Speciﬁc requirements-
Part 11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY)
Speciﬁcations. IEEE 802.11g-2003.

[8] IEEE Standard for Information technology-Telecommunications and information ex-
change between systems-Local and metropolitan area networks-Speciﬁc requirements-
Part 11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY)
Speciﬁcations. IEEE 802.11n-2009.

[9] IEEE Standard for Information technology-Telecommunications and information ex-
change between systems-Local and metropolitan area networks-Speciﬁc requirements-
Part 11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY)
Speciﬁcations. IEEE 802.11ac.

[10] IEEE Standard for Information technology-Telecommunications and information ex-
change between systems-Local and metropolitan area networks-Speciﬁc requirements-
Part 11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY)
Speciﬁcations. IEEE 802.11ad.

[11] Moscow state

university

video

quality measurement

tool

(msu

vqmt),

http://compression.ru/.

[12] Qualcomm research, https://www.qualcomm.com/invention/research.

[13] Streaming media, http://www.streamingmedia.com/.

[14] x264 - a free h.264/avc encoder, http://www.videolan.org/developers/x264.html.

74

[15] x264 licensing , https://licensing.x264.org/en/.

[16] Ieee standard for telecommunications and information exchange between systems -
lan/man speciﬁc requirements - part 11: Wireless medium access control (mac) and
physical layer (phy) speciﬁcations: High speed physical layer in the 5 ghz band. IEEE
Std 802.11a-1999, pages 1–102, Dec 1999.

[17] Ieee standard for information technology– local and metropolitan area networks– speciﬁc
requirements– part 11: Wireless lan medium access control (mac) and physical layer
(phy) speciﬁcations amendment 6: Wireless access in vehicular environments. IEEE Std
802.11p-2010 (Amendment to IEEE Std 802.11-2007 as amended by IEEE Std 802.11k-
2008, IEEE Std 802.11r-2008, IEEE Std 802.11y-2008, IEEE Std 802.11n-2009, and
IEEE Std 802.11w-2009), pages 1–51, July 2010.

[18] A. Schulman, D. Levin, and N. Spring. Crawdad data set umd/sigcomm2008 (v. 2009-

03-02).

[19] Siripuram Aditya and Sachin Katti. Flexcast: Graceful wireless video streaming. In
Proceedings of the 17th Annual International Conference on Mobile Computing and
Networking, MobiCom ’11, pages 277–288, New York, NY, USA, 2011. ACM.

[20] R. Ahola, A. Aktas, J. Wilson, K. R. Rao, F. Jonsson, I. Hyyrylainen, A. Brolin,
T. Hakala, A. Friman, T. Makiniemi, J. Hanze, M. Sanden, D. Wallner, Yuxin Guo,
T. Lagerstam, L. Noguer, T. Knuuttila, P. Olofsson, and M. Ismail. A single-chip
cmos transceiver for 802.11a/b/g wireless lans. IEEE Journal of Solid-State Circuits,
39(12):2250–2258, Dec 2004.

[21] Fan Bai, Daniel D. Stancil, and Hariharan Krishnan. Toward understanding character-
istics of dedicated short range communications (dsrc) from a perspective of vehicular
network engineers. In Proceedings of the Sixteenth Annual International Conference on
Mobile Computing and Networking, MobiCom ’10, pages 329–340, New York, NY, USA,
2010. ACM.

[22] Laura Bernad´o, Nicolai Czink, Thomas Zemen, and Pavle Belanovic. Physical layer
simulation results for ieee 802.11 p using vehicular non-stationary channel model. In
Communications Workshops (ICC), 2010 IEEE International Conference on, pages 1–5.
IEEE, 2010.

[23] Apurv Bhartia, Yi-Chao Chen, Swati Rallapalli, and Lili Qiu. Harnessing frequency
diversity in wi-ﬁ networks. In Proceedings of the 17th Annual International Conference
on Mobile Computing and Networking, MobiCom ’11, pages 253–264, New York, NY,
USA, 2011. ACM.

[24] Bastian Bloessl, Mario Gerla, and Falko Dressler. Ieee802. 11p in fast fading scenar-
In Proceedings of the

from traces to comparative studies of receive algorithms.

ios:

75

First ACM International Workshop on Smart, Autonomous, and Connected Vehicular
Systems and Services, pages 1–5. ACM, 2016.

[25] Bastian Bloessl, Michele Segata, Christoph Sommer, and Falko Dressler. An ieee 802.11
a/g/p ofdm receiver for gnu radio. In Proceedings of the second workshop on Software
radio implementation forum, pages 9–16. ACM, 2013.

[26] Bastian Bloessl, Michele Segata, Christoph Sommer, and Falko Dressler. Towards an
open source ieee 802.11 p stack: a full sdr-based transceiver in gnu radio. In Vehicular
Networking Conference (VNC), 2013 IEEE, pages 143–149. IEEE, 2013.

[27] C. Phillips and S. Singh. Crawdad data set pdx/vwave (v. 2009-07-04).

[28] B. Chen, Z. Zhou, Y. Zhao, and H. Yu. Eﬃcient error estimating coding: Feasibility

and applications. IEEE/ACM Transactions on Networking, 20(1):29–44, Feb 2012.

[29] Tim Das. Practical considerations for low noise ampliﬁer design. Freescale Semiconduc-

tors, Inc, 2013.

[30] Thierry Ernst. The information technology era of the vehicular industry. SIGCOMM

Comput. Commun. Rev., 36(2):49–52, April 2006.

[31] Matthew Gast. 802.11 wireless networks: the deﬁnitive guide. O’Reilly Media, Inc.,

2005.

[32] F. Gringoli

and

L. Nava.

Open

ﬁrmware

for WiFi

networks.

http://www.ing.unibs.it/openfwwf/.

[33] Daniel Halperin, Wenjun Hu, Anmol Sheth, and David Wetherall. Tool release: Gath-
ering 802.11n traces with channel state information. SIGCOMM Comput. Commun.
Rev., 41(1):53–53, January 2011.

[34] Bo Han, Lusheng Ji, Seungjoon Lee, B. Bhattacharjee, and R.R. Miller. All bits are
not equal - a study of ieee 802.11 communication bit errors. In INFOCOM 2009, IEEE,
pages 1602–1610, April 2009.

[35] Bo Han, Lusheng Ji, Seungjoon Lee, Bobby Bhattacharjee, and Robert R Miller. Are all
bits equal?: experimental study of ieee 802.11 communication bit errors. IEEE/ACM
Transactions on Networking (TON), 20(6):1695–1706, 2012.

[36] Bo Han, Aaron Schulman, Francesco Gringoli, Neil Spring, Bobby Bhattacharjee,
Lorenzo Nava, Lusheng Ji, Seungjoon Lee, and Robert Miller. Maranello: Practical
partial packet recovery for 802.11. In Proceedings of the 7th USENIX Conference on
Networked Systems Design and Implementation, NSDI’10, pages 14–14, Berkeley, CA,
USA, 2010. USENIX Association.

76

[37] Wei Hu, Jin Xie, and Zhenghao Zhang. pmore: Exploiting partial packets in oppor-
tunistic routing. In 2011 IEEE Global Telecommunications Conference-GLOBECOM
2011.

[38] Jun Huang, Yu Wang, and Guoliang Xing. Lead: Leveraging protocol signatures for
improving wireless link performance. In Proceeding of the 11th Annual International
Conference on Mobile Systems, Applications, and Services, MobiSys ’13, pages 333–
346, New York, NY, USA, 2013. ACM.

[39] Szymon Jakubczak, Hariharan Rahul, and Dina Katabi. One-Size-Fits-All Wireless
In Proc. Eighth ACM SIGCOMM HotNets Workshop, New York City, NY,

Video.
October 2009.

[40] Kyle Jamieson and Hari Balakrishnan. Ppr: Partial packet recovery for wireless net-
works. In Proceedings of the 2007 Conference on Applications, Technologies, Architec-
tures, and Protocols for Computer Communications, SIGCOMM ’07, pages 409–420,
New York, NY, USA, 2007. ACM.

[41] Jirka Klaue, Berthold Rathke, and Adam Wolisz. Evalvid a framework for video
transmission and quality evaluation. In Peter Kemper and WilliamH. Sanders, editors,
Computer Performance Evaluation. Modelling Techniques and Tools, volume 2794 of
Lecture Notes in Computer Science, pages 255–272. Springer Berlin Heidelberg, 2003.

[42] Zoran Kotevski and Pece Mitrevski.

ICT Innovations 2009, chapter Experimental
Comparison of PSNR and SSIM Metrics for Video Quality Estimation, pages 357–366.
Springer Berlin Heidelberg, Berlin, Heidelberg, 2010.

[43] V. Kukshya and H. Krishnan. Experimental measurements and modeling for vehicle-
to-vehicle dedicated short range communication (dsrc) wireless channels. In IEEE Ve-
hicular Technology Conference, pages 1–5, Sept 2006.

[44] Zhenjiang Li, Yaxiong Xie, Mo Li, and Kyle Jamieson. Recitation: Rehearsing wireless
packet reception in software. In Serge Fdida, Giovanni Pau, Sneha Kumar Kasera, and
Heather Zheng, editors, MobiCom, pages 291–303. ACM, 2015.

[45] Kate Ching-Ju Lin, Nate Kushman, and Dina Katabi. Ziptx: Harnessing partial packets
in 802.11 networks. In Proceedings of the 14th ACM International Conference on Mobile
Computing and Networking, MobiCom ’08, pages 351–362, New York, NY, USA, 2008.
ACM.

[46] Xiao Lin Liu, Wenjun Hu, Qifan Pu, Feng Wu, and Yongguang Zhang. Parcast: Soft
video delivery in mimo-ofdm wlans. In Proceedings of the 18th Annual International
Conference on Mobile Computing and Networking, Mobicom ’12, pages 233–244, New
York, NY, USA, 2012. ACM.

77

[47] Daniele Lo Iacono and Teo Cupaiuolo. Power eﬃcient sdr implementation of ieee

802.11a/p physical layer. J. Signal Process. Syst., 73(3):281–289, December 2013.

[48] Ratul Mahajan, Jitendra Padhye, Sharad Agarwal, and Brian Zill. High performance ve-
hicular connectivity with opportunistic erasure coding. In Presented as part of the 2012
USENIX Annual Technical Conference (USENIX ATC 12), pages 237–248, Boston,
MA, 2012. USENIX.

[49] J. Mittag, S. Papanastasiou, H. Hartenstein, and E. G. Strom. Enabling accurate cross-
layer phy/mac/net simulation studies of vehicular communication networks. Proceedings
of the IEEE, 99(7):1311–1326, July 2011.

[50] Allen Miu, Hari Balakrishnan, and Can Emre Koksal. Improving loss resilience with
multi-radio diversity in wireless networks. In Proceedings of the 11th Annual Interna-
tional Conference on Mobile Computing and Networking, MobiCom ’05, pages 16–30,
New York, NY, USA, 2005. ACM.

[51] Jens-Rainer Ohm. Bildsignalverarbeitung fuer multimedia-systeme. skript, 1999. Skript.

[52] T. H. Pham, I. V. McLoughlin, and S. A. Fahmy. Shaping spectral leakage for ieee
802.11p vehicular communications. In 2014 IEEE 79th Vehicular Technology Conference
(VTC Spring), pages 1–5, May 2014.

[53] I.E. Richardson. H.264 and MPEG-4 Video Compression: Video Coding for Next-

generation Multimedia. Wiley, 2004.

[54] A. Sassi, F. Charﬁ, L. KAMOUN, Y. Elhillali, and A. Rivenq. OFDM transmission
International Journal of Computer

performance evaluation in V2X communication.
Science Issues, 9:141–148, 2012.

[55] Sayandeep Sen, Syed Gilani, Shreesha Srinath, Stephen Schmitt, and Suman Banerjee.
Design and implementation of an “approximate” communication system for wireless
media applications. SIGCOMM Comput. Commun. Rev., 41(4):–, August 2010.

[56] Souvik Sen, Naveen Santhapuri, Romit Roy Choudhury, and Srihari Nelakuditi. Ac-
curate: Constellation based rate estimation in wireless networks. In Proceedings of the
7th USENIX Conference on Networked Systems Design and Implementation, NSDI’10,
pages 12–12, Berkeley, CA, USA, 2010. USENIX Association.

[57] S. Subik and C. Wietfeld.

Integrated pmr-broadband-ip network for secure realtime
multimedia information sharing. In Technologies for Homeland Security (HST), 2011
IEEE International Conference on, pages 20–25, Nov 2011.

78

[58] W. Vandenberghe, I. Moerman, and P. Demeester. Approximation of the ieee 802.11p
standard using commercial oﬀ-the-shelf ieee 802.11a hardware. In 2011 11th Interna-
tional Conference on ITS Telecommunications, pages 21–26, Aug 2011.

[59] Mythili Vutukuru, Hari Balakrishnan, and Kyle Jamieson. Cross-layer wireless bit rate

adaptation. SIGCOMM Comput. Commun. Rev., 39(4):3–14, August 2009.

[60] G. Wang, S. Zhang, K. Wu, Q. Zhang, and L. M. Ni. Tim: Fine-grained rate adaptation

in wlans. IEEE Transactions on Mobile Computing, 15(3):748–761, March 2016.

[61] K Daniel Wong. Fundamentals of wireless communication engineering technologies,

volume 98. John Wiley & Sons, 2011.

[62] Grace R. Woo, Pouya Kheradpour, Dawei Shen, and Dina Katabi. Beyond the bits:
Cooperative packet recovery using physical layer information.
In Proceedings of the
13th Annual ACM International Conference on Mobile Computing and Networking,
MobiCom ’07, pages 147–158, New York, NY, USA, 2007. ACM.

[63] Xiufeng Xie, Xinyu Zhang, Swarun Kumar, and Li Erran Li. pistream: Physical layer
informed adaptive video streaming over lte. GetMobile: Mobile Comp. and Comm.,
20(2):31–34, October 2016.

[64] Yaxiong Xie, Zhenjiang Li, and Mo Li. Precise power delay proﬁling with commodity
wiﬁ. In Proceedings of the 21st Annual International Conference on Mobile Computing
and Networking, MobiCom ’15, pages 53–64, New York, NY, USA, 2015. ACM.

[65] M. Zargari, L. Y. Nathawad, H. Samavati, S. S. Mehta, A. Kheirkhahi, P. Chen,
K. Gong, B. Vakili-Amini, J. A. Hwang, S. W. M. Chen, M. Terrovitis, B. J. Kaczynski,
S. Limotyrakis, M. P. Mack, H. Gan, M. Lee, R. T. Chang, H. Dogan, S. Abdollahi-
Alibeik, B. Baytekin, K. Onodera, S. Mendis, A. Chang, Y. Rajavi, S. H. M. Jen, D. K.
Su, and B. A. Wooley. A dual-band cmos mimo radio soc for ieee 802.11n wireless lan.
IEEE Journal of Solid-State Circuits, 43(12):2882–2895, Dec 2008.

[66] X. Zhao, H. Lu, C. W. Chen, and J. Wu. Adaptive hybrid digital-analog video transmis-
sion in wireless fading channel. IEEE Transactions on Circuits and Systems for Video
Technology, 26(6):1117–1130, June 2016.

79