Full Text 01

Automatic control in TCP over wireless
NIELS M
OLLER
Licentiate Thesis
Stockholm, Sweden, 2005
TRITA-S3-REG-0504
ISSN-1404-2150
ISBN-91-7178-143-9
KTH
SE-100 44 Stockholm
SWEDEN
Akademisk avhandling som med tillstand av Kungl Tekniska hogskolan framlagges
till oentlig granskning for avlaggande av licentiatexamen i telekommunikation
2005-09-30 i D34.
c Niels Moller, 2005
Tryck: Universitetsservice US AB
iii
Abstract
Over the last decade, both the Internet and mobile telephony have become parts of
daily life, changing the ways in which we communicate and search for information. These
two distinct technologies are now slowly merging. The topic of this thesis is tcp over
wireless, and the automatic control that is used within the system, from the link-layer
power control to the end-to-end congestion control. The thesis consists of three main
contributions.
The rst contribution is a proposed split-connection scheme for downloads to a mobile
terminal. A wireless mobile terminal requests a le or a web page from a proxy, which in
turn requests the data from a server on the Internet. During the le transfer, the radio
network controller (rnc) sends radio network feedback (rnf) messages to the proxy. These
messages include information about bandwidth changes over the radio channel, and the
current rnc queue length. A novel control mechanism in the proxy uses this information
to adjust the sending rate. The stability and convergence speed of the proxy controller is
analyzed theoretically. The performance of the proposed controller is compared to end-
to-end tcp Reno, using ns-2 simulations for some realistic scenarios. It is shown that the
proxy controller is able to reduce the response time experienced by users, and increase the
utilization of the radio channel. The changes are localized to the rnc and the proxy; no
changes are required to the tcp implementation in the terminal or the server.
The second contribution is the analysis of an uplink channel using power control and
link-layer retransmissions. To be able to design the link-layer mechanisms in a systematic
way, good models for the link-layer processes, and their interaction with tcp, are essential.
The use of link-layer retransmissions transforms a link with constant delay and random
losses into a link with random delay and almost no losses. As seen from the tcp endpoints,
the dierence between such a link and a wired one is no longer the loss rate, but the packet
delay distribution. Models for the power control and link-layer retransmissions on the link
are used to derive the packet delay distribution, and its impact on tcp performance is
investigated.
The nal contribution involves ways to optimize the link-layer processes. The main
result is that tcp performance, over a wireless link with random retransmission delays, can
be improved by adding carefully chosen articial delays to certain packets. The articial
delays are optimized o-line and applied on-line. The additional delay that is applied to
a packet depends only on the retransmission delay experienced by that same packet, and
this information is available locally at the link.
Preface
Natverk, natforband (Lat. opus reticulatum), bygnk., ett murforband,
hvari stenarna aro stallda pa horn, sa att fogarna bli diagonala, (se g.),
stundom anvandt i den gamla romerska och nagon gang i den fornkrist-
na byggnadskonsten.

Afven en natliknande ornering kallas natverk. Jfr
Mur, sp. 1372.
G. H. W Upmark, Nordisk familjebok, 1914.
This thesis is the result of research in the area where computer networking and
the theory of automatic control intersect. The project Towards a self-regulating
Internet at kth was started in 2002. This project is a collaboration between the
automatic control group at the department of Signals, Sensors, and Systems at
kth, and the communications networking group, which at the time
1
was part of
the department of Microelectronics and Information Technology at kth. I was the
rst graduate student in automatic control, to be involved in this project.
In recent years, applying control theory to computer networking, as an emerging
research eld, has attracted interest from three dierent communities and cultures:
the control theory community, the networking community, and the telecommunica-
tion community.
When I got involved in the project, I was well versed in mathematics and com-
puter science, and I knew the basics of control theory. I had considerable practical
experience with computer networking, from the period I spent helping building the
rst dorm network in Linkoping, and from programming various network applica-
tions, but I was not familiar with networking theories and research. The traditional
telecom world was, and still is, to some degree, an alien world.
So it has been interesting and rewarding to try to get the dierent traditions and
viewpoints to t together. For a slightly exaggerated example, say that a mobile
and a base station exchange packets, and that packets that are lost due to noise
on the radio are retransmitted (this is called link-layer retransmissions, and is one
central issue in this thesis). A control theorist would think about the base station
and mobile as a feedback control system, and draw a simplied block diagram. A
computer network researcher might view the retransmissions as a dubious hack,
1
Today, the organization is dierent. Both groups are now part of the same department, and
have the fortune of being located in the same building.
v
vi
intended to make the important end-to-end protocol tcp work better, and then
implement it in the ns-2 network simulator. A telecommunications researcher
might view the retransmissions as a quality of service parameter, which the operator
could charge extra for. Since the work has been done within the automatic control
group, I have tried to primarily keep a control perspective sometimes translating
networking concepts into the language of automatic control.
The research described in this thesis has been partially funded by the Swedish
Research Council, by the Swedish Foundation for Strategic Research, by the Eu-
ropean Commission through the hycon and runes projects, and by the Swedish
Center for Internet Technologies.
Photos are kindly provided by Kjell Enblom and Todd Klassy. The latter photo
is distributed under the Creative Commons attribution license.
I would like to thank my advisor Karl Henrik Johansson, who has read and
commented on my draft papers more times than anybody else, and Carlo Fischione,
who has helped me to get the radio communications issues right. For the radio
network feedback work, thanks goes to Ines Cabrera Molero, who performed the
simulations and wrote the needed ns-2 code, as part of an excellent master of science
thesis, and also to the people at Ericsson who were involved:

Ake Arvidsson, Justus
Petersson, and Robert Skog.
I would like to thank my fellow PhD students and the other friendly people in
the control group. Finally, I would like to thank Cecilia for her encouragement,
ever since I rst started on this project.
Contents
Contents vii
1 Introduction 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 IP networking and the end-to-end principle . . . . . . . . . . . . . . 1
1.3 The TCP protocol . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.4 Wireless links . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.5 TCP over wireless . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.6 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2 Background 9
2.1 Congestion control in TCP . . . . . . . . . . . . . . . . . . . . . . . 9
2.2 Wireless links . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3 TCP over wireless . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3 Radio network feedback 21
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.2 Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.3 Control structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.4 Stability analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.5 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.6 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
4 TCP over a power controlled channel 37
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.2 Radio model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.3 Link-Layer Retransmissions . . . . . . . . . . . . . . . . . . . . . . . 44
4.4 TCP/IP Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.5 Throughput Degradation . . . . . . . . . . . . . . . . . . . . . . . . 49
4.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
5 Improving the link layer 55
vii
viii CONTENTS
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
5.2 Retransmission as feedback . . . . . . . . . . . . . . . . . . . . . . . 55
5.3 IP packet delay . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
5.4 Improving the link-layer . . . . . . . . . . . . . . . . . . . . . . . . . 58
5.5 Robustness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
5.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
6 Conclusions and further work 65
6.1 Radio network feedback . . . . . . . . . . . . . . . . . . . . . . . . . 65
6.2 Delay variations of wireless links . . . . . . . . . . . . . . . . . . . . 66
6.3 Approaches to TCP over wireless . . . . . . . . . . . . . . . . . . . . 67
6.4 Towards a systematic design for layer separation . . . . . . . . . . . 68
Bibliography 71
A Proof of Lemma 1 75
Chapter 1
Introduction
The main theme of this thesis is the use of tcp/ ip over wireless links,
and interactions between processes in dierent layers.
1.1 Motivation
Over the last decade, both the Internet and mobile telephony have become parts
of daily life, changing the ways we communicate and search for information. These
two distinct tools are now slowly merging, both at the surface, and in the underlying
communication infrastructure.
A wireless mobile Internet means that mobile gadgets are rst-class citizens of
the Internet. Any communication task, e.g., email, web browsing, Internet radio,
or peer-to-peer le sharing, that is possible with a stationary computer connected
to the Internet, should be equally possible with a suitable mobile gadget.
Enabling a wireless mobile Internet is a huge task. One prerequisite is that
tcp/ip, the two most important protocols on the Internet, must work satisfactorily
across a heterogeneous network consisting of an assortment of stationary and mobile
devices, connected by dierent types of wired and wireless links.
The focus of this thesis is on the tcp protocol, the problems encountered when
using tcp over cellular radio links, and the various feedback control loops used
within the system. This rst chapter provides an introduction to the tcp-over-
wireless problem, and outlines the contributions in this thesis. Chapter 2 provides
additional background material and references. The main contributions are found in
Chapters 35. Finally, Chapter 6 discusses the results, and outlines some directions
for future research.
1.2 IP networking and the end-to-end principle
Layered design of communication systems is a modularization technique, where
each layer at a particular node needs to know how to communicate with the layers
1
2 CHAPTER 1. INTRODUCTION
Application layer
Presentation layer
Session layer
Transport layer
Network layer
Link layer
Physical layer
Application
Transport
ip
Link
Physical media
Figure 1.1: Left: The osi networking stack. Right: The ip networking model.
directly above and below at the local node, but only to the same layer at remote
nodes. E.g., the Internet Protocol (ip) layer (or networking layer) needs to know
how to transmit ip packets using the local link-layer, but it does not need to know
anything about the receivers link-layer.
The Open Systems Interconnection (osi) reference model species seven layers,
illustrated to the left in Figure 1.1. From bottom to top: physical layer, link layer,
networking layer, transport layer, session layer, presentation layer and application
layer. ip networks uses a somewhat simpler model. At the core, we have the ip
layer, corresponding to the networking layer of the osi model. The ip layer is a
fairly primitive packet transport service, which provides unreliable best eort packet
delivery between nodes, identied by their ip addresses (32 bits for ip version 4,
128 bits for ip version 6). Packets may be dropped, duplicated or delivered out of
order.
The power of ip, the Inter-network protocol, is that it is used to communicate
across heterogeneous networking environments, e.g., ethernet, point-to-point links,
and cellular networks. These dierent technologies are accommodated as the ab-
stract notion of a link and a link layer. The link layer is directly below the ip layer
in the networking stack, and it provides ip packet transport between nodes that
share the same link. There are multitude of dierent link types in the Internet,
and a multitude of transport layer and application layer protocols, but they are all
used with a single packet transport service: ip. This is illustrated to the right in
Figure 1.1.
Above the ip layer, we have the transport layer, where the Transmission Control
1.3. THE TCP PROTOCOL 3
Protocol (tcp) is the protocol of primary interest. tcp is responsible for dividing
a data stream into packets, ensure reliable delivery even when the ip layer loses,
reorders, or duplicates packets, and at the same time it senses the state of the
network to avoid overloading it. In the context of ip networking (as opposed to
the construction of applications and application protocols), everything above the
transport layer is usually referred to as application layer, with no subdivision
into session layer, presentation layer, etc.
There are no sophisticated mechanisms for resource reservation or allocation
built in to the ip layer; the normal response for a router or link that is overloaded
is to simply discard the packets it cannot handle, and leave to the communicating
endpoints to sort things out best they can. This is an important design choice:
The network core is simple, while endpoints must be quite sophisticated in order
to work well together with the network. This design is known as the end-to-end
principle.
1.3 The TCP protocol
The dominating transport layer protocol is tcp. It is used for all kinds of data
streams: long-lived low bandwidth interactive trac, e.g., telnet and ssh sessions,
short-lived le transfers, e.g., web trac, longer lived bulk transfers, e.g., ftp and
le sharing, and almost real-time trac, e.g., Internet radio.
A tcp connection is a bidirectional, ow controlled, reliable stream of data
between two endpoints, identied by ip-address and port number. Our primary
interest is in tcp connections for transfer of smaller or larger les. For this tcp us-
age, it is desirable to get as much data as possible through the network, while at the
same time we must avoid overloading the network, and share available bandwidth
in a fair way with other users.
tcp uses a sliding window ow control. The window limits the amount of data
that can be sent without waiting for acknowledgement (ack) from the receiver.
When the window is constant, this results in the so called ack clock; the timing
of each sent packet is determined by the reception of the ack for an earlier packet.
One can think about the sliding window and the ack clock as a peculiar inner
control loop which determines the sending rate; when the roundtrip time uctuates,
the sliding window gives an average sending rate of one full window per average
roundtrip time. The window size is adjusted depending on received acks, and it is
the details of this outer loop that dier between tcp variants.
The objective of the tcp window control is to get a high throughput, close
to the connections fair share of the available bandwidth, and at the same time
avoid overloading the network. The fair share can vary due to varying amounts of
competing cross trac, and also due to network changes such as to routing updates
or radio links with time-varying capacity.
Figure 1.2: Left: Point-to-point microwave link, which connected the student dorm
network in Linkoping to the the campus network and the Internet, 19931996.
Photo by Kjell Enblom, RydNet. Right: Cell phone tower in Oregon. Photo by
Todd Klassy.
1.4 Wireless links
Until the 1990s, the Internet consisted almost exclusively of wired links, consisting
of copper cables and optical bers. The occasional point-to-point microwave links
were usually quite reliable high power devices using parabolic antennas, such as the
link at the left of Figure 1.2.
Today, mobile wireless communication is common place, thanks to the infras-
tructure of mobile telephony. Radio communication in the cellular telephony system
is more dicult than a point-to-point microwave link, due to lower transmission
power, mobility, and often lack of a direct line-of-sight between transmitter and
receiver. The mobile telephony system is tailored to handle these diculties of the
available radio channel.
For the future, we need to integrate the mobile telephony system with the
Internet infrastructure. To make the mobile terminal a rst class citizen of the
Internet, we need to use tcp and ip all the way out to the terminal. To make
such an integrated system work eciently is challenging, because the wireless link
available to the terminal has dierent characteristics than the wired links that tcp
was originally designed for.
The operation of a wireless link is much more complex than traditional electrical
or optical links. The various mechanisms used, and the corresponding input and
output signals, are shown in Figure 1.3.
1.4. WIRELESS LINKS 5
Encode + Decode
Scheduling Redundancy Rate Power sir Error
g n
Figure 1.3: Wireless link mechanisms. From left to right: Buer of frames to send,
channel coder, modulator, transmission amplier, transmission antenna, channel
gain, channel noise, reception antenna, demodulator, and channel decoder.
On the sending side, at the left in the gure, we have a buer of frames to
send, controlled by a scheduler which decides which frame to transmit in which
time slot. Next, a channel coder, adding redundancy for error correction coding,
and a modulator, selecting the modulation scheme and the particular waveform to
use. Finally, a power amplier and the transmission antenna.
The disturbances on the radio channel can usually be accurately modelled as a
multiplicative gain disturbance and an additive noise, g and n in the gure. These
are time-varying due to movement of antennas and obstacles, and due to varying
level of background noise and interference.
On the right, we have the receiver, consisting of a demodulator that detects
the used waveform, and a decoder that tries to correct symbol errors. Besides the
actual data, the output signals from the receiver are the signal to interference ratio
(sir)
1
and a per frame indication saying if the frame could be recovered after error
correction. The available control signals are the transmission power, the sending
rate used by the modulator, the amount of redundancy added by the channel coder,
and the scheduling, i.e., decisions on which frame to transmit when, as well as
possibly deciding to stay silent when conditions are too bad.
The gure shows a single wireless link. In practice, we have multiple links oper-
ating in the same area and frequency range. In this thesis, the focus is on cellular
architecture, where base stations are distributed geographically, each base station
handles a number of mobile terminals, and the access scheme used is wideband code
division multiple access (wcdma).
In a cellular system, the design tradeos are quite dierent for the uplink, i.e.,
transmission from a mobile terminal to the base station, and the downlink, i.e.,
transmissions from the base station to the mobile terminals. In the uplink, capacity
is limited by the interference between mobile terminals, which makes power control
crucial. The performance of the power control inuences not only the battery life of
the terminals, but also the overall capacity and stability of the system. The power
used by one terminal appears as an additive disturbance for the other users. In
1
The sir compares the power of the signal to the power of the interference from other users.
This is dierent from the signal to noise ratio, which compares the power of the signal to the power
of the thermal noise. In cellular systems, interference usually dominates over thermal noise.
Internet
End-to-end
Split connection Link-layer
Figure 1.4: Approaches to the tcp over wireless problem.
the downlink, on the other hand, the base station can use scheduling or orthogonal
modulation to limit the interference between transmissions to individual terminals.
Instead, the downlink capacity is limited by the transmission power of the base
station, and by interference from neighboring base stations.
The various transmission parameters, e.g., rate, redundancy, scheduling and
power, are controlled using cascaded feedback loops, working on dierent timescales.
Power control based on the received sir is the fastest loop, working on a 1 ms time
scale, scheduling works on a timescale of time slots, 10 ms, while adaptation of
the rate and of the amount of redundancy is typically another order of magnitude
slower.
1.5 TCP over wireless
When sending ip packets over a wireless link, the ip layer sees a link where one or
more of the capacity, the loss rate, and the delay characteristics varies with time.
Use of the available link-layer control mechanisms (power control, rate adaptation,
forward error correction, and retransmission scheduling) enables us to make trade-
os between capacity, loss and delay, but it is not possible to achieve constant high
capacity, low delay, and low loss, which is characteristic of a wired link.
The tcp congestion control mechanisms were designed to adapt to the type
of disturbances that are common in a wired network, where available bandwidth
varies due to cross trac and occasional routing or capacity changes, and delays
are caused by constant propagation delays and queueing delays.
When tcp is used over the cellular infrastructure, the result is often that both
end-to-end throughput and radio link utilization are quite poor. This is because
the dynamic properties of tcp and of wireless links do not t well together.
The objective of work on the tcp-over-wireless problem is to achieve both good
end-to-end throughput and ecient utilization of the radio resources, preferably
with as small changes to existing infrastructure and protocols as possible. The wild
fauna of proposed solutions to the tcp-over-wireless problem can be classied as
follows:
1.6. CONTRIBUTIONS 7
End-to-end: Improve the tcp-protocol to adapt better to the disturbances from
wireless links.
Split-connection: Introduce a proxy in the wired network, close to the wireless
link. The proxy acts as an endpoint for a tcp connection over the wired
network, and communicates with the wireless terminal using a specialized tcp
version over the wireless part of the network, or even using some completely
dierent protocol.
Link-layer: Design link-layer tradeos that results in link properties that plain
tcp handles better.
These approaches are illustrated in Figure 1.4.
The proxy mechanism described in Chapter 3 falls into the split-connection class.
It uses an explicit http proxy. Another variant of the split-connection approach
(which is not discussed in detail in this thesis) is called performance enhancing
proxies. Such a proxy does not act as a tcp endpoint, instead, it can insert,
delay or delete packets in the stream, usually acknowledgements from the mobile
terminal [7]. By doing so, the proxy can hide some of the wireless disturbances
from the endpoint on the wired network, or force the end point to react dierently
to events on the wireless link.
Chapter 5 follows the link-layer approach. It describes a way to change the
packet delay distribution by adding some additional delay, which reduces the prob-
ability of spurious timeout and improves tcp throughput.
It is likely that mechanisms from more than one of these classes should be used.
An important aspect of the problem is to understand how the needed disturbance
rejection work should be divided between end-to-end mechanisms and mechanisms
closer to the wireless link.
1.6 Contributions
This thesis contains three fairly independent contributions.
Chapter 3 proposes a split-connection proxy solution for tcp over the high-
speed downlink packet access (hsdpa) channel in wcdma. The connection
between the proxy and the terminal does not use standard tcp congestion
control. Instead the proxy uses feedback control and feedforward control,
based on explicit messaging from the radio network controller (rnc), which
has detailed knowledge of the current radio conditions. Simulations and quan-
titative results have been presented as
I. Cabrera Molero, N. Moller, J. Petersson, R. Skog,

A. Arvidsson,
O. Flardh, and K. H. Johansson. Cross-layer adaptation for tcp-
based applications in wcdma systems. In IST Mobile & Wireless
Communications Summit, Dresden, 2005.
The stability analysis will be presented as
N. Moller, I. Cabrera Molero, K. H. Johansson, J. Petersson, R. Skog,
and

A. Arvidsson. Using radio network feedback to improve tcp
performance over cellular networks. In IEEE CDC-ECC, Seville,
2005.
The ns-2 simulations for this chapter were done by Ines Cabrera Molero, as
part of her Master Thesis project [33].
Chapter 4 considers the uplink in a cellular system with power control and
local retransmissions. Models for all layers from the power control up to tcp
allow quantitative prediction of tcp performance degradation. This work is
the chronologically rst, and has been presented as
N. Moller and K. H. Johansson. Inuence of power control and
link-level retransmissions on wireless tcp. In Quality of Future In-
ternet Services, volume 2811 of Lecture Notes in Computer Science.
Springer-Verlag, Stockholm, 2003.
K. Jacobsson, N. Moller, K. H. Johansson, and H. Hjalmarsson.
Some Modeling and Estimation Issues in Control of Heterogeneous
Networks. In Mathematical Theory of Networks and Systems, Leu-
ven, 2004.
N. Moller, C. Fischione, K. H. Johansson, F. Santucci, and F. Graziosi.
Modeling and control of ip transport in cellular radio links. In ifac
World Congress, Prague, 2005.
Chapter 5 describes a way to make a wireless link with a discrete delay dis-
tribution more friendly to tcp, by adding articial delays to certain packets.
This work has been presented as
N. Moller, K. H. Johansson, and H. Hjalmarsson. Making retrans-
mission delays in wireless links friendlier to tcp. In IEEE CDC,
Bahamas, 2004.
and also at MTNS 2004 above.
Chapter 2
Background
tcp over wireless is a complex system with several cascaded feedback
loops. This chapter describes the use of automatic control in each sub-
system.
2.1 Congestion control in TCP
The objective of congestion control is to keep the load of the network close to the
available capacity, and at the same time share the available capacity fairly between
ows. Fairness is a peripheral issue in this thesis; but it must be noted that fairness
is an important constraint in the design of transport protocols for the Internet.
The tcp protocol was developed in the late 1970s, resulting in the Internet Stan-
dard rfc 793 [15]. The principles for tcp congestion control were developed a few
years later, in response to experience of congestion collapses in the Internet [24].
Window-based control
The most important concept in tcp congestion control is that of the congestion
window. The window is the amount of data that has been sent, but for which no
acknowledgement has yet been received. A constant congestion window means that
one new packet is transmitted for each ack that is received.
The sending rate is controlled indirectly by adjusting the congestion size. The
standard way of doing this is documented in rfc 2581 [2], usually referred to
as tcp Reno. It is described in this section. Two other common variants are
tcp NewReno [18] and tcp with selective acknowledgements (sack) [31, 19]. Before
explaining the control mechanisms, we have to look into how tcp detects packet
losses.
9
10 CHAPTER 2. BACKGROUND
Acknowledgements and loss detection
At the receiving end, acknowledgement packets are sent in response to received data
packets. tcp uses cumulative acknowledgements: Each acknowledgement includes
a sequence number that says that all packets up to that one has been received.
Equivalently, the acknowledgement identies the next packet that the receiver ex-
pects to see.
When packets are received out of order, each received packet results in an ac-
knowledgement, but they will identify the largest sequence number such that all
packets up to that number has been received. E.g., if packets 1, 2, 4, and 5 are
received, four acknowledgements are generated. The rst says I got packet #1,
I expect packet #2 next, while the next three acknowledgements all say I got
packet #2, I expect packet #3 next. The last two acknowledgements are duplicate
acks, since they are identical to some earlier ack.
On the sending side, there are several possible reasons why duplicate acks are
received: Packets delivered by the network out of order, packets dropped by the
network, and ack packets duplicated by the network.
Packet losses are detected by the sender in two ways:
Timeout. If a packet is transmitted and no ack for that packet is received
within the retransmission timeout interval (rto), the packet is considered
lost.
Fast retransmit. If three duplicate acks are received, the next expected
packet from these acks is considered lost. Note that this can not happen if
the congestion window is smaller than four packets.
Packets that are lost, as detected by either of these mechanisms, are retransmitted.
Furthermore, congestion control actions are also based on these loss signals, as
described below.
The value for rto is not constant, but based on measured average and variation
of the rtt. It is also modied by the exponential backo mechanism.
TCP congestion control state
There are four distinctive states in the tcp congestion control, illustrated in Fig-
ure 2.1, and two state variables related to congestion control: The congestion win-
dow cwnd and the slow start threshold ssthresh. Typical initial values when tcp
leaves the idle state and enters the slow start state are a cwnd of 2 packets, and
a ssthresh that is the maximum value allowed by the wire protocol and by the
receiving end.
We look at the operation of each of the four states in turn.
2.1. CONGESTION CONTROL IN TCP 11
Idle Slow start
Congestion
avoidance
Exponential
backo
Fast
recovery
Figure 2.1: tcp state diagram. Transitions back to the idle state are omitted.
Slow start
The slow start state is the rst state entered when a ow is created, or when a
ow is reactivated after being idle. The slow start state can also be entered as
the result of a timeout. In this state, cwnd is increased by one packet for each
non-duplicate ack. The eect is that for each received ack, two new packets are
transmitted. This implies that the congestion window, and also the sending rate,
increases exponentially, doubling once per rtt.
It may seem strange to refer to an exponential increase of the sending rate as
slow start; the reason is that in the early days, tcp used a large window from the
start, and the introduction of the slow start mechanism did slow down connection
startup.
Slow start continues until either
cwnd > ssthresh, in which case tcp enters the congestion avoidance state, or
a timeout occurs, in which case tcp enters the exponential backo state, or
three duplicate acks are received, in which case tcp enters the fast recovery
state.
The motivation for the slow start state is that when a new ow enters the
network, and there is a bottleneck link along the path, then the old ows sharing
that link need some time to react and slow down before there is room for the new
ow to send at full speed.
Congestion avoidance
In congestion avoidance mode, cwnd is increased by one packet per rtt (if cwnd
reaches the maximum value, it stays there). This corresponds to a linear increase
in the sending rate. On timeout, tcp enters the exponential backo state, and on
three duplicate acks, it enters the fast recovery state.
The motivation for this congestion avoidance mechanism is that since tcp does
not know the available capacity, it has to probe the network to see at how high a
rate data can get through. Aggressive probing would make the system unstable,
and a single packet increase seems to work well in practice.
Exponential backo
tcp enters the exponential backo mode after timeout. Several actions are taken
when entering this state:
The lost packet is transmitted.
The state variables are updated by ssthresh cwnd/2, cwnd 1 packet.
The rto value is doubled.
If the retransmission timer expires again with no ack for the retransmitted packet,
the packet is repeatedly retransmitted, rto is doubled, and ssthresh is set to
1 packet [16]. The upper bound for the rto is on the order of one or a few minutes.
Exponential backo continues until an acknowledgement for the packet is re-
ceived, in which case tcp enters the slow start phase, or the tcp stack or application
gives up and closes the connection.
The motivation for the exponential backo mechanism is that timeouts, in par-
ticular repeated timeouts, are a sign of severe network congestion. In order to avoid
congestion collapse, the load on the network must be decreased considerably and
repeatedly, until it reaches a level with a reasonably small packet loss probability.
Fast recovery
tcp enters the fast recovery state after it detects three duplicate acks. When
entering this mode, the rst actions of tcp is to retransmit the lost packet, and set
ssthresh cwnd/2.
tcp then continues to send new data at approximately the same rate, one new
packet of data for each received duplicate ack. In rfc 2581 [2], this is described
using a fairly complex procedure that articially inates cwnd.
If no ack for the retransmitted packet is received within the rto interval, tcp
enters the exponential backo state. Otherwise, when an ack for the retransmitted
packet is nally received, tcp sets cwnd = ssthresh, i.e., half the cwnd value at the
start of the recovery procedure, and enters the congestion avoidance state.
2.1. CONGESTION CONTROL IN TCP 13
If more than one packet is lost within the same window, fast recovery is limited
in that it can recover only one packet per rtt. This is the main problem addressed
by tcp NewReno and tcp sack.
The motivation for the fast recovery mechanism is that the reception of duplicate
acks indicates that the network is able to deliver new data to the receiver. Hence,
the network is not severely congested, and we can keep inserting new packets into
the network at the same rate as packets are delivered, at least for a while.
On the other hand, the loss of a packet also indicates that the network is on the
border of congestion. At the end of the fast recovery procedure, cwnd is halved.
tcp restarts the probing of the congestion avoidance state at a lower sending rate,
at which it did not get any losses.
It should also be noted that halving the cwnd also implies that tcp will stay
silent for about half an rtt, waiting for acks that reduce the number of outstanding
packets, until the outstanding packets match the new window size.
Control theoretic view of congestion control
In control theoretic terms, a mechanism for congestion control of a ow in the
network must somehow estimate the ows fair share of the available bandwidth,
and then use a sending rate based on this estimate. The main diculties in doing
this is that:
Information about the network state, available at the endpoints, is scarce.
The value being estimated, the fair share of available capacity, is time-varying
and subject to disturbances.
Sending rates based on an overestimation of the true value will quickly lead
to queues building up in the network, leading to both packet losses and longer
roundtrip times for those packets that are not lost.
A router in the network can be expected to know the capacities of attached
links, the arrival rate for recent trac, and the lengths of its queues. But, in the
spirit of the end-to-end principle, routers should not be expected or required to
maintain any per ow information. The signalling from the network infrastructure
to tcp end nodes is severely limited. The signals available to end nodes are the
arrival and timing of acks, and, if supported by the network, a single bit of explicit
congestion notication (ecn) attached to each ack [37].
Disturbance sources include varying levels of tcp cross trac, varying levels
of non-congestion controlled cross trac, routing and topology changes, and, for
wireless links, capacity changes due to variations of the radio channel.
The ACK clock
When tcps congestion window is kept constant, the sender transmits one new
packet for each received ack, which is referred to as the ack clock. The number
of packets inside the network (be they data packets or acks) is kept constant. The
ack clock is based on the idea that by controlling the number of packets inside the
network, we can control the load of the network.
. . . the packet ow is what a physicist would call conservative: A new
packet isnt put into the network until an old packet leaves. The physics
of ow predicts that systems with this property should be robust in the
face of congestion.
Van Jacobson, [24]
The average transmission rate is one window of data per roundtrip time. In tcp
congestion control, the control signal is the window size, and the actual sending rate
is controlled only indirectly by adjusting the window.
Additive increase, multiplicative decrease
The combination of the congestion avoidance and fast recovery mechanisms, often
called Additive Increase, Multiplicative Decrease (aimd), leads to a tcp sending
rate that is sawtooth shaped. When a small number of tcp ows share a bottle-
neck link, the senders tend to synchronize: All the ows increase their sending
rate until the trac exceeds capacity and the bottleneck queue starts to build up.
When the queue overows, packets from all ows are dropped, forcing the ows to
halve their window sizes almost simultaneously, and then the process starts over.
However, when a large number of ows share a bottleneck, synchronization seems
not to happen [3].
Bottleneck queues and RTT variations
The ack clock partially addresses the problem of overestimation leading to queues
building up, since if a queue on the path builds up, the roundtrip time increases,
which leads to a decreased sending rate. The ack clock by itself is however not a
very satisfactory control mechanism.
The reason is that in the Internet, there is a quite sharp distinction between
bottleneck and non-bottleneck queues. A queue that is not a bottleneck will natu-
rally stay almost empty. On the other hand, the congestion avoidance mechanism
of tcp will cause the bottleneck queues to stay close to overow, at least if there
are a large number of ows. Then, the roundtrip time will be a constant value that
depends only on which queues are bottlenecks, plus a small random noise.
2.2 Wireless links
Wireless links result in disturbances to the ip packet transport that are dierent
from the typical disturbances on wired links. The disturbances on a wireless link
may include
2.2. WIRELESS LINKS 15
Encode + Decode
Scheduling Redundancy Rate Power sir Error
g n
Figure 2.2: Wireless link mechanisms. From left to right: Buer of frames to send,
channel coder, modulator, transmission amplier, transmission antenna, channel
gain, channel noise, reception antenna, demodulator, and channel decoder.
Packet losses. Unlike wired networks, a wireless link can have a high packet
loss rate, and packet losses that are not associated with network congestion.
Capacity variation. Due to, e.g., movements of the wireless terminal or ob-
stacles, the wireless link may switch between lower and higher sending rates.
Delay variation. Due to, e.g., local retransmissions, the variation of packet
delay may be qualitatively dierent from that of a wired link.
Temporary outages. As an extreme form of delay variation, packets may be
buered for a long time, i.e., an rtt or more, during a temporary outage of
the radio channel, and be transmitted only after the radio channel is available
again.
Automatic control mechanisms of the wireless mechanisms at the link, see Fig-
ure 2.2, can be used to reject some disturbances, or trade one type of residual
disturbance for another. The signals and the corresponding feedback loops work
on dierent time scales.
Power control
Power control is the fastest loop. sir-based power control uses feedback from the
measured sir at the receiver to control the transmission power at the sender. This
is referred to as the power control inner loop, and the control objective is to keep
the received sir close to a reference value sir
ref
.
Assume we use a constant sending rate and a constant amount of redundancy
for forward error correction (fec). Let sir
bad
denote a sir at which reception is
barely possible, corresponding to, say, a frame loss probability of 50%.
If the inner loop uses a reference value sir
ref
> sir
bad
, and the fading, i.e.,
variations of the channel gain, is slow enough, then the inner loop by itself can
reject the fading disturbance, and ensure that the received sir is higher than sir
bad
almost all of the time. But when the fading is faster other strategies are needed.
If the fading is too fast for the inner loop, then the variation of the channel gain
g is larger than the variation of the power, controlled by the inner loop. The result
is that the received sir will vary in the same way as the channel gain, and at the
times when g is smallest, we can have sir < sir
bad
.
To deal with this situation, the target value for the sir can be increased. Then
the varying received SIR is essentially increased by a constant. This shortens or
eliminates the time intervals when sir < sir
bad
. Furthermore, if these time intervals
are made short enough, fec with a suitable amount of redundancy can be used to
recover from any demodulation errors.
In Chapter 4 we will consider an outer power control loop that adjusts sir
ref
using feedback from the frame error signal from the decoder at the far right in
Figure 2.2.
The achievable performance of feedback power control is limited by two main
factors: One is the feedback delay, which makes it impossible to track fast changes
in the channel gain, as discussed above. The other is limitations on the control
signal. Increasing the transmission power causes interference for other users, which
coupled to the other users power control may jeopardize stability of the system: If
every user increases his own transmission power without limit, all communication is
drowned in interference. Global stability is one of the central issues in distributed
power control systems [21], in particular for the uplink in cellular systems.
Forward error correction
Forward error correction adds redundant parity bits to each frame before trans-
mission. On reception, a limited number of errors in the data or parity bits can
be corrected. fec does not rely on feedback, and is therefore not limited by the
feedback delay. It can correct errors due to short disturbances of duration down
to the transmission time for a single bit. The limit to the eectiveness of fec is
instead in terms of the number of damaged bits per frame.
One technique that is commonly used to increase the eectiveness of fec over
channels with bursts of errors is interleaving: Parity bits are added to each frame,
and then transmission of frames are interleaved in time. Then, an error burst
corresponding to the transmission time of, say, 100 bits, will not appear as 100 bit
errors in a single frame, but as 10 bit errors in each of 10 frames.
fec is particularly attractive for wireless channels subject to multipath fading.
Such channels exhibit sharp dips in the channel gain, which it is not feasible to
reject by increasing the transmission power. But since the channel is in such a dip
only a fairly small proportion of the time, fec with suitably chosen parameters can
correct the resulting errors. To get the most out of a channel, power control and
fec should be tuned together [6].
Scheduling and link-layer retransmissions
Scheduling is the decision of which frame to send when. For a shared downlink
channel, such as the hsdpa channel in the universal mobile telecommunications
system (umts), the base station is free to decide which user to transmit for in each
2.3. TCP OVER WIRELESS 17
time slot. The decision can be based on measured or predicted channel quality for
the dierent users.
For an uplink dedicated channel, such as the one considered in Chapter 4, it is
not so much a question of when to use the channel, but a question of which frame
to send in each time slot. In particular, whether or not a new data frame should
be transmitted, or an old frame be retransmitted.
Tuning of the parameters for the link-layer control, i.e., power control, fec, and
rate adaptation, usually tries to optimize the channel throughput, under constraints
such as the total cost of base stations and other infrastructure
1
. The selected
parameters may correspond to a fairly high frame loss rate, up to 10% [13, 32].
The impact of such a high frame loss rate, on the quality of the services that
the link is used for, naturally depends on the type of the service in question. As
discussed in the next section, for any service that depends tcp ows, a small loss
rate is essential for high quality.
By retransmitting damaged frames, one buys a small loss rate, and pays with
increased, random, delays. General recommendations on the use of local retransmis-
sions on Internet links are documented in rfc 3366 [17]. Link-layer retransmissions
can be viewed as a feedback loop, driven by the frame error signal. Chapter 4 mod-
els one link retransmission scheme, and analyzes the impact on tcp throughput.
2.3 TCP over wireless
Poor tcp performance over wireless links is a well-known problem. The traditional
explanation for poor tcp performance is that the wireless link drops packets due
to noise and fading on the radio channel, and that tcp interprets all packet losses
as indications of network congestion [14]. An overview of the problems and of
proposed mechanism at the link or transport layer can be found in [41].
Since tcp performance problems have been observed also on links with link-
layer retransmissions, and hence few lost packets, it is not sucient to understand
how tcp is aected by packet losses. Using link-layer retransmissions trades losses
for delay, and the resulting delays can also inuence tcp negatively. In Chapter 4,
we investigate the delay distribution for the uplink in a cellular system, and its
interaction with tcp.
Understanding and improving tcp behavior over wireless links is a topic of
intense research. This section describes a few representative contributions from
each of the main approaches, but it does not attempt to be an exhaustive survey
of the eld.
1
This makes the most sense when operators are charging per received bit. With a at rate
business, maybe the tradeos would be dierent?
End-to-end
The end-to-end approach intends to improve tcp behavior over fairly general classes
of wireless links.
The Eifel algorithm adds extra information to packets (using the standard tcp
timestamp option) to make it possible for the sender to distinguish between ac-
knowledgements for original transmissions and for retransmissions [29]. This infor-
mation makes it possible for the tcp sender to react more intelligently to certain
types of disturbances. In particular, when all packets are delayed for a time longer
than the tcp timeout value but not lost (typical for a short temporary outage in
a link employing local retransmissions), recovery is improved signicantly by using
the Eifel algorithm.
The idea of tcp Westwood is to estimate the available bandwidth over the
path, based on the timing of received acks [30]. When recovering after a loss, this
estimate is used to set the new rate to a less pessimistic value than the rate used
by standard tcp.
For wireless links where losses are the primary problem, one natural approach is
to try to detect if a packet loss is due to congestion or to transmission errors [10, 39,
20]. In general, the end-to-end approach leads to interesting estimation problems,
where end nodes use the limited data available, such as ack timing, to extract
information about the network state.
Link-layer
The link-layer approach uses models for the underlying radio, and uses these models
for evaluation, tuning, and design, of link-layer mechanisms. In the context of tcp
over wireless, common quality measures are user response time and tcp throughput.
Within this class, tcp over links with random errors and no retransmissions is
analyzed in [1]. In [27], tcp throughput over a link is simulated, for various radio
conditions and link-layer retransmission schemes.
There are many important link tradeos that have been investigated, including
tcp downlink performance in a wcdma system with joint rate and power adapta-
tion [22], the tradeo between link-layer fec and tcp throughput [12, 5], and the
tradeos between fec, arq and transmission power [11, 6].
The Gilbert-Elliot model is widely used in the literature; this is a two-state
Markov model with one good and one bad state for the radio channel. tcp
behavior over links that use local retransmissions on top of a Gilbert-Elliot radio
channel is analyzed in [8, 36],
Cross-layer mechanisms
The end-to-end principle implies that the network and transport layers in a node
can not or should not know anything about the link layers in remote parts of
the network (the network layer in a node naturally has to be aware about the links
2.3. TCP OVER WIRELESS 19
that are physically attached to the same node). Architecturally, this is a very sound
design, but in some circumstances, in particular when wireless links are involved,
it may lead to suboptimal performance.
It is possible to improve the performance by introducing explicit cross-layer
signalling. This term denotes signalling between layers that are separated both
geographically and in networking stack order. Typically, the signalling is between
the transport layer in one or both endpoints, and the link layer attached to an
intermediate radio link. There are at least two types of cross-layer signalling. The
link can inform the transport endpoint of the radio link state, such as the current
capacity. One example of such radio network feedback is considered in Chapter 3.
It is also possible to let a transport endpoint inform the radio link layer of its
requirements, e.g., the preferred tradeo between loss rate and delay.
Besides the general observation that a controller can usually do a better job
the more information it has about the process being controlled, another important
reason why cross-layer design is considered, is based on overall deployment issues.
Deploying a new version of tcp is a complex and time consuming process. Stan-
dardization is fairly slow, there are a large number of tcp-implementations that
must be updated, and there are huge number of devices that are attached to the
Internet, which must all be updated or replaced until a new version of tcp can
replace the current version.
For a solution to the tcp-over-wireless problem to be feasible in the short term,
i.e., deployable within a few years, it is essential that it does not require that a
majority of Internet devices be upgraded. It is preferable if upgrades are isolated
to a certain class of devices, e.g., Internet enabled mobile phones, or to a subset of
the network within a single administrative domain, e.g., one operators network.
Using cross-layer mechanisms provides some additional freedom in the design,
which can make a huge dierence for the deployability of a solution. Whether or not
cross-layer mechanisms are fundamentally needed for a high performance wireless
Internet is an open and controversial question [9, 26].
Chapter 3
Radio network feedback
In this chapter, we investigate a cross-layer approach to improve the
performance for web browsing (or other forms of tcp download) over
a cellular network.
3.1 Introduction
When a user downloads a le or a web page from the Internet to a mobile terminal,
the path through the network includes a mobile terminal, a radio link to an op-
erators cellular network, a radio network controller (rnc), the operators internal
core network, a gateway between the operators network and the Internet, and a
web server attached to the Internet. Using a proxy in this way, as illustrated in
Figure 3.1, is an example of the split-connection approach to tcp over wireless.
Let us consider the ow of data from the web server to the mobile terminal, using
a high bandwidth wireless channel, such as the high-speed downlink packet access
(hsdpa) channel in wcdma. The bandwidth over the radio link varies with time,
including outage periods when no communication is possible. Using a pure end-
to-end protocol such as tcp, any information about the state of the radio link has
to pass down to the terminal and back, adding a signicant delay. Furthermore,
during a temporary outage, no signalling from the terminal can reach the other
endpoint.
This motivates a cross-layer approach: The rnc is well informed about the
radio link state, and it is well connected to the Internet. So we can let the rnc
transmit radio network feedback (rnf) messages to the sender. A straight-forward
application of this idea requires that the tcp stack at the web server is modied
to take advantage of the rnf-signalling, which is unrealistic, at least in the short
term.
Instead, we use a http proxy in the operators network.
1
Virtually all web
browsers support http proxying, and the proxy communicates with the target web
1
Using a more transparent tcp proxy is also possible.
21
22 CHAPTER 3. RADIO NETWORK FEEDBACK
rnc Proxy
rnf message Variable bandwidth
tcp tcp
Figure 3.1: Proposed architecture. The mobile terminal on the left downloads a le
from the server on the right, via the proxy. During the transfer, the rnc generates
rnf messages including information about the current bandwidth over the radio
link, and the current rnc queue length. The proxy uses this information to adjust
its sending rate.
server using standard http and tcp. Thus, only the rnc and the proxy, both part
of the operators network, need to be aware of the rnf mechanism.
This chapter proposes and analyzes a new control mechanism based on rnf.
The performance improvement, compared to end-to-end tcp, is evaluated in ns-2
simulations. It turns out that the rnf mechanism yields signicant improvements
for both the radio link utilization and the end-to-end response times experienced
by the user. Sections 3.2 and 3.3 describe the system architecture and control
structure. The proposed controller is described and analyzed in Section 3.4, and
the discrete time implementation is described in Section 3.5. Finally, in Section 3.6,
the controller is evaluated using ns-2 simulations in three realistic use cases.
3.2 Architecture
Consider the transfer of a le, from a server on the Internet, to a mobile terminal
attached via an operators radio access network. Within the operators network,
there are two special nodes: An rnc, which among other things is responsible for
allocating bandwidth to user connections, and a proxy, which acts as a gateway
between the operators network and the Internet. The endpoints use standard tcp
to communicate with the proxy. The proxy, on the other hand, adapts its sending
rate towards the terminal using a custom control algorithm, which is aided by extra
information provided by the rnc. We make the reasonable assumptions that the
bottleneck for the connection is the radio channel, and that the operators network,
between proxy and rnc, does not suer congestion.
The architecture is illustrated in Figure 3.1, where two tcp connections have
been established: one between the terminal and the proxy and one between the
proxy and the server. The rnc sends rnf messages to the proxy, including infor-
mation about the current bandwidth allocation and the queue length in the rnc.
The rnf messages are sent every time the bandwidth changes, and also periodically
with a relatively long period time, e.g., one second.
We will compare this system to the nominal setup, in which there is a direct
3.3. CONTROL STRUCTURE 23
Proxy
controller
w 3G
network
q
b
q
ref
Figure 3.2: Control structure. The controller uses event-triggered feedforward of
the available radio link bandwidth b, and time-triggered feedback of the rnc queue
length q.
tcp connection between the terminal and the server.
We use the following notation. The bandwidth of the radio link is b.
f
is the
time it takes for a packet that is sent by the proxy to reach the rnc, and
b
is the
time it takes for a packet that is forwarded by the rnc to reach the terminal, and for
the corresponding acknowledgement to get back to the proxy. The corresponding
rtt of the cellular network, excluding queueing delay at the rnc, is =
f
+
b
, and
the pipe size, or bandwidth-delay product, is b. We assume that b,
f
, and
b
are constant most of the time, with occasional step changes. The queue length at
the rnc is denoted q, so that the rtt is + q/b. The current window size at the
proxy is denoted w.
3.3 Control structure
The objective is to achieve a high utilization of the radio link and maintain the rnc
queue close to a reasonably small reference value q
ref
, by controlling the sending
rate of the proxy. Like in standard tcp, we use a window-based algorithm, but
unlike standard tcp, we take advantage of explicit information provided by the
rnc.
The control signal is the non-negative window size w, which indirectly deter-
mines the sending rate. The network provides the proxy controller with two kinds
of information. The control structure, with feedforward of radio bandwidth b, and
feedback of queue length q, is shown in Figure 3.2.
When the available bandwidth over the radio channel is changed, the rnc sends
an rnf message to the proxy to inform it about the new bandwidth. The controller
uses an event-triggered feedforward mechanism that takes advantage of this infor-
mation, and resets the window size to a new value,
w =

b + q
ref
(3.1)
Here is an estimate of the propagation delay , and

b is the new bandwidth from
the rncs rnf message.
Between bandwidth updates, the rnc periodically sends rnf messages with the
current queue length. The controller compares this feedback information to the
reference value and adjusts its window,
w
k+1
= w
k
+ c(q
ref
q
k+1
) (3.2)
The feedback loop uses a quite long sampling time, on the order of one second, and
it is designed to compensate for the bias caused by a feedforward controller based
on uncertain information. On shorter time scales, the transmission window is xed,
and transmission is governed by the usual ack clock.
3.4 Stability analysis
To investigate the stability of the control mechanisms, we use a ow level model.
The system input signal is w(t) and the output signal is q(t). The sending rate is
related to the window size by
r(t) =
w(t)
+ q(t
b
)/b
where +q(t
b
)/b is the rtt corresponding to the queue length at the time the
most recently acknowledged packet was forwarded by the rnc. The rate of change
of the queue (as long as it is neither full nor empty) equals the dierence between
the received rate and the bandwidth,
q(t) = r(t
f
) b =
w(t
f
)
+ q(t )/b
b
=
w(t
f
) b q(t )
+ q(t )/b
(3.3)
Feedforward control
With feedforward control, the sender uses estimates

b and to compute a con-
stant control signal according to (3.1). Substituting this expression into the queue
dynamics (3.3), gives
q(t) =
b + q
ref
b q(t )
+ q(t )/b
This time-delayed dierential equation has an equilibrium point q
= q
ref
+
b b,
and can be rewritten as
q(t) =
q(t
f
) q
+ q(t
f
)/b
Dene
= + q
/b as the rtt corresponding to the equilibrium. The linearized

dynamics are
q(t) =
1
(q(t ) q
)
3.4. STABILITY ANALYSIS 25
The Nyquist criterion gives asymptotic stability since
< 1 <

2
To estimate the convergence time, consider the poles of the linearized system, i.e.,
the solutions to
s +
e
s
= 0
Since the system is stable, all solutions lie in the half plane Re s < 0. Note that
for any such pole, it holds that |s| = e
Re s
/
> 1/
, so the poles lie outside

of the disk of radius 1/
around the origin. Let us derive a bound > 0 for

the stability exponent of the system, i.e., a bound such that Re s < for all
solutions to the characteristic equation above. The trajectories will locally converge
to the equilibrium faster than e
t
, and 1/ is an upper bound for the system time
constant. We use the following result.
Lemma 1 The equation x = e
x
, with > 0, has a solution x > 0 if and only if
e. Furthermore, the solutions satisfy 1/( 1) < x < 2.
Proof : See Appendix A.
First consider a real pool, s = x. Then x = e
x
/
. With =
/, the rst
part of the lemma yields
e, and the second part gives x > 1/(
/ 1), or
x > 1/(
). So in this case, we have the bound > 1/(
).
Next, consider a complex pole in the left half plane, s = x +iy, with x, y > 0.
Again we want to nd a bound > 0 such that all poles are to the left of Re s = .
So assume that x . Denote z = e
s
= e
x
(cos y i sin y), and assume that
(x iy) =
s = z, so that s is a pole of the system. First, note that
y |
s| = |z| = e
x
e
Require to be such that e
< /2, which is always possible for some > 0.

Then cos is decreasing on the interval (y, /2). We get for the real part
>
x = Re(
s) = Re z = e
x
cos y > cos
e
This implies that > (1/
) cos(e
). So if we choose small enough, this

inequality is violated and then there can be no poles with x . Hence, dene
= {smallest positive solution , to
= cos
e
}
We obtain the general bound for the stability exponent as
= min
_
1
_
1 2 3 4 5 6 7 8 9 10
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
/
Figure 3.3: Stability exponent for = 1. The upper curve is the derived
conservative estimate = min(1/(
),
). The lower curve is computed

numerically using an order seven Pade approximation for the delay.
Note that depends only on the ratio
/. Figure 3.3 shows as a function

of
/. The convergence speed peeks for
= e, which is when the system has a

real double pole at 1/. It is easy to obtain the bound > 1/(4) for the interval
<
< e. To summarize, we have the following result.

Proposition 1 The feedforward control results in an asymptotically stable system,
with an equilibrium, which corresponds to a deviation from the reference value q
ref
that equals the estimation error for the pipe size,

b b. As long as
< e, the
convergence time constant for the linearized system is smaller than 4.
Feedback control
Next, consider feedback control of the window size, using queue length samples. The
primary objective of the feedback control is to cancel the bias that results when
using feedforward control from estimates

b and that are uncertain. A natural
approach is to use an integrating controller,
w(t) = c
_
q(t
b
) q
ref
_
(3.4)
With this control law, the closed loop system is described by the equations
_
_
_
q(t) =
w(t
f
) b q(t )
+ q(t )/b
w(t) = c
_
q(t
b
) q
ref
_
(3.5)
3.4. STABILITY ANALYSIS 27
Since q and w are non-negative, these equations are valid only in the interior of the
domain q, w 0. On the border w = 0 we have instead
w(t) = max
_
0, c
_
q(t
b
) q
ref
_
_
and similarly for q on the line q = 0.
The equations have an equilibrium point given by q
= q
ref
and w
= b + q
ref
.
The rtt in stationarity is
= + q
ref
/b. With this notation, the system can be
written as
_
_
_
q(t) =
w(t
f
) w
(q(t ) q
+ (q(t ) q
)/b
w(t) = c
(q(t
b
) q
)
Note that the queue delay inuences the system dynamics via the gain in the
expression for q, but the queue is not in the signaling path. In the rnf architecture,
the rnf messages say how long the queue is, but the rnf messages themselves do
not suer any queueing delay. In other words, the time delays in the dierential
equations are constant and not depending on the system state.
In the following analysis, we ignore the propagation delay =
f
+
b
. We
will show that in this case the system is globally asymptotically stable. Introduce
q = q q
and w = ww
. Then, the non-linear system without propagation delay

can be written as
_
q(t) =
w q
+ q/b
w(t) = c
q
We make a change of variables by introducing a virtual time s, dened by dt =
(
+ q/b)ds. The virtual time corresponds to the number of rtts. Then

_
_
d q
ds
= w q
d w
ds
= c
q
c
b
q
2
These dierential equations are still valid only in the interior of the region w, q 0,
non-negativity has to be enforced at the borders. The phase portrait of this system
(for unit parameters) is shown in Figure 3.4. Our goal is to show that this system is
globally asymptotically stable in the region w, q 0. Dene a candidate Lyapunov
function
V ( q, w) = q
2
( q + A) +
Bb
2c
w
2
q = (w 1) (q 1)
w = ( (q 1) (q 1)
2
) (q<1|w>0)

0 0.5 1 1.5 2 2.5 3
0
0.5
1
1.5
2
2.5
3
q
w
Figure 3.4: Phase portrait of q-w dynamics for the feedback control system.
where A and B are constants to be dened. Then
dV
ds
= q(3 q + 2A)( w q)
+ B w q(b
q)
= q
2
(3 q + 2A)
+ q w(2ABb
)
+ q
2
w(3 B)
Fix B = 3, A = 3b
/2. Note that

q + A = q q
ref
+ 3(b + q
ref
)/2
= q + q
ref
/2 + 3b/2 > 0
3 q + 2A = 3(q q
ref
) + 3(b + q
ref
)
= 3(q + b) > 0
Hence,
V ( q, w) = q
2
_
q +
q
ref
+ 3b
2
_
+
3b
2c
w
2
dV
ds
= 3 q
2
(q + b) 0
3.5. IMPLEMENTATION 29
By LaSalles Theorem, the trajectory converges to an invariant subset of the line
q = 0. The only such set is the origin ( q, w) = (0, 0), corresponding to (q, w) =
(q
, w
), cf., the point (1, 1) in Figure 3.4. We have thus proved the following result:
Proposition 2 The feedback control system, with negligible propagation delay in
the cellular network, is globally asymptotically stable.
3.5 Implementation
In practice, the sender does not have access to the continuous-time function q(t),
but rely on samples sent from the rnc. Denote the sampling time by T. For the
ow model to be valid, an averaging over at least one rtt is needed, so T should
not be shorter than the rtt. We next consider the case of a T several times larger
than the rtt. (In the simulations presented in next section, the sampling time is
one second.)
Between samples, the controller keeps w constant. The queue dynamics will
converge to the stationary value within about twice the system time constant.
From the analysis for a constant window size, in Section 3.4, we know that the
stationary value is q
= w b. Hence, if T is suciently large, and w(t) = w

k
for
kT < t (k + 1)T, then
q
k+1
= q
_
(k + 1)T
_
w
k
b
The continuous-time integrator controller corresponds to the discrete-time con-
troller
w
k+1
= w
k
+ c(q
ref
q
k+1
)
It turns out that c = 1 is a good choice for the controller gain, because then
q
k+2
w
k+1
b w
k
+ q
ref
(w
k
b) b = q
ref
The queue will thus converge to the target value within two samples, 2T.
3.6 Results
The proposed proxy controller takes advantage of rnf messages and thus gets bet-
ter performance out of the link compared to standard tcp. The gain has three
main components: Connection startup, outage recovery, and bandwidth adapta-
tion. They are discussed in detail below. After that, the rnc queue transients are
examined, and at end of the section, we describe a quantitative study for three
realistic use cases.
Several advantages with the proxy controller is illustrated by the ns-2 simula-
tions shown in Figure 3.5, which depicts available bandwidth (dotted), sending rate
for proxy controller (dashed), and sending rate for nominal end-to-end tcp (solid).
In this gure, note that the transmission via the proxy is completed in about 51 s,
0 10 20 30 40 50 60 70
0
200
400
600
800
1000
1200
1400
1600
Figure 3.5: Available bandwidth (dotted), sending rate for proposed proxy con-
troller (dashed), and sending rate for nominal end-to-end tcp (solid). The link
under-utilization is visible as the area between the bandwidth and the sending rate
curves. The sending rates in this and the other gures in this chapter are in kbit/s,
averaged over 300 ms. The time axis is in seconds. (That the sending rates some-
times seem to exceed the available bandwidth is an artifact of the sampling of the
ns-2 simulation.)
while the end-to-end tcp transmission lasts approximately 10 s longer. It is clear
that the proxy controller is able to utilize the available bandwidth much better
than the nominal end-to-end controller.
Connection startup
The connection startup phase is illustrated in the very beginning of Figure 3.5. A
zoom in is shown in Figure 3.6. The initial delay before the sending rate starts to rise
is the rtt between terminal and server. The rate curve for nominal tcp is governed
by the slow start mechanism of tcp (Section 2.1). The ow from the proxy to the
terminal does not use slow start at all. However, the ow from the server to the
proxy does use slow start, which explains why the rate curve for the proxy can not
jump instantaneously to 1 Mbit/s. The reason that the proxy solution outperforms
nominal tcp, even though both curves are directly or indirectly limited by the slow
3.6. RESULTS 31
1 1.5 2 2.5 3 3.5
0
100
200
300
400
500
600
700
800
900
1000
Figure 3.6: Confection startup. The response for the proxy controller is limited
by the slow start of the serverproxy connection. It is still much faster than the
response of end-to-end tcp.
start mechanism, is that the the rtt between the proxy and the server is much
shorter than the rtt between the terminal and the server.
Outage recovery
When an outage occurs, tcp goes into timeout and retransmits lost packets re-
peatedly using exponential backo. It will not notice that the outage has ended
until it receives an ack for one of the retransmitted packets, which in the worst
case can be 60 seconds after the outage ended. This is illustrated in Figure 3.7,
which zoom in the second section (after slow start) of Figure 3.5. Furthermore,
when the outage has ended, tcp goes into congestion avoidance mode, where the
sending rate is increased only linearly, which means that it takes quite some time
until the sending rate approaches the available bandwidth. The proxy solution does
not suer from this additional delay, since it gets explicit information from the rnc
when the outage starts (
b = 0) and ends (
b > 0). After the outage ends, the proxy

immediately starts sending at the right rate.
5 10 15 20 25
0
100
200
300
400
500
600
700
800
900
1000
Figure 3.7: Outage recovery. Since the proxy controller does not use exponential
backo or congestion avoidance, it recovers much faster than the nominal tcp from
an outage.
Bandwidth adaptation
When the bandwidth of the radio channel is increased, a tcp sender in congestion
avoidance mode increases its sending rate slowly. When the bandwidth is decreased,
the tcp sender will not notice until some packet is lost due to buer overow in
the rnc. (In our simulations this does not happen, since we have a very large rnc
buer.) Thanks to the feedforward mechanism in the proxy controller, it reacts
more quickly, as illustrated in Figure 3.8, which is a zoom in of the middle section
of Figure 3.5. The performance gain is not as signicant as the gains in the slow
start and outage recovery phases.
RNC queue transient
The size of the rnc queue is time-varying. When the bandwidth of the radio channel
is changed, the queue will initially drift o from the reference value, cf., the analysis
in Section 3.4. The magnitude of the variation depends on the propagation delay
3.6. RESULTS 33
28 30 32 34 36 38 40 42 44
1000
1050
1100
1150
1200
1250
1300
1350
1400
1450
1500
Figure 3.8: Bandwidth adaptation. The proxy controller reacts faster to radio
bandwidth changes than the nominal tcp controller.
for the rnf messages. A typical value of the propagation delay is 50 ms, but can
depend on the implementation as well as dynamic behavior. Figure 3.9 shows the
evolution of the queue length when the bandwidth is changed. It indicates what is
a reasonable size of the rnc buer.
Quantitative evaluation
In order to evaluate the performance of the proposed proxy controller in some
practical situations, a set of use cases was dened and a ns-2 simulation study
was performed for each of them. The proxy setup in Figure 3.1 is compared to
a nominal setup, in which there is no proxy but only a direct tcp connection
between the terminal and the server. The links between the server and the rnc are
set to 2 Mbit/s. The wireless link between the rnc and the terminal has a lower
time-varying bandwidth and a one-way delay of 60 ms. We also use an average rtt
(server-terminal-server) of 300 ms. The slow start threshold was set to 50, to ensure
that the tcp sender does not enter the congestion avoidance phase prematurely.
Figure 3.5 illustrates the transmission of a 5.7 Mbyte le.
The performance of the proxy setup was evaluated in terms of the average link
utilization and the time to serve a user (ttsu). The ttsu is dened as the time
27 27.2 27.4 27.6 27.8 28 28.2 28.4 28.6 28.8 29
1
2
3
4
5
6
7
Figure 3.9: Evolution of the queue length, when the bandwidth is increased at 28 s
in the simulation. The target value for the queue length is 6 packets.
elapsed from when the connection is established (syn packet sent by the requester)
until the last data packet is received at the terminal. We consider three use cases,
aimed at modeling common services to be provided over a wcdma system. Use
case 1 models a 4 Mbyte le transfer. Use case 2 imitates web browsing and is based
on statistical data [35, 23]. In this use case, three les (whose sizes vary between
51.5 Kbyte and 368.5 Kbyte) are downloaded to a terminal during one web session.
Use case 3 models a large (25 Mbyte) le transfer. The results are summarized in
Table 3.1. The ttsu is particularly improved with the proxy controller for relatively
small le downloads (such as web browsing). Link utilization is close to 100% for
the proxy controller for large les.
3.7. SUMMARY 35
ttsu [s] Utilization [%]
Use case 1 Nominal 47.3 77%
Proxy 39.2 98%
Proxy 3.0 78%
Proxy 220.5 98%
Table 3.1: Comparison of time to serve a user (ttsu) and link utilization for three
use cases.
3.7 Summary
tcp-based applications in a wcdma system were studied in two setups: a nominal
one that employs end-to-end tcp Reno and a new one with an intermediate proxy.
The proxy solution uses rnf messages from the data-link layer in the rnc to the
transport layer in the proxy. The proxy uses feedforward control to adjust to a
varying bandwidth, and feedback control to maintain the rnc queue length close to
a desired value. It was shown that the resulting closed-loop system is stable and that
the convergence time is linked only to the propagation delay in the cellular system.
The performance gain was estimated using ns-2 simulations. In these evaluations,
we considered three realistic use cases. The proxy controller was able to reduce the
time to serve users and increase the radio link utilization. The proposed control
architecture might substantially improve the user experience of wireless Internet.
The proposal is transparent to both endpoints, so no modications need to be done
to the terminal or the server.
Chapter 4
TCP over a power controlled
channel
We develop a model for a cellular link, including processes from the con-
trol of the transmission power up to the end-to-end congestion control.
4.1 Introduction
When studying tcp over wireless, the model for the channel loss process is of central
importance. Since it is dicult to accurately describe all factors inuencing the
radio transmission, analysis of the wireless system has to be based on models that
are simplied. When using a model for control design, even crude models often
work remarkably well, but it is of course necessary that the model captures the
range of dynamics we wish to control.
In the literature, it is common to focus on the frame loss process, which is often
described using the Gilbert-Elliot model. The Gilbert-Elliot model is a two-state
Markov model, corresponding to one good and one bad state for the radio channel.
Each state is associated with a constant loss probability.
In this chapter, we depart from the Gilbert-Elliot model, by looking at a realistic
model of a power controlled channel [40]. We assume that the channel uses an inner
loop power control that is able to keep the signal to interference ratio (sir) close
to a reference value, viewing the residual control error as noise. The loss process of
the power controlled channel is generated by the outer loop power control, which
adjusts the reference value based on experienced losses.
When deploying umts, which uses power controlled channels, the target value
for the resulting frame loss rate can be on the order of several percents. The
wireless link employs local retransmissions of damaged frames. These link-layer
retransmissions transform a link with constant delay and random losses into a link
with random delay and almost no losses.
tcp was designed for wired networks, where packet losses are usually due to
37
38 CHAPTER 4. TCP OVER A POWER CONTROLLED CHANNEL
pc
sir
ref
+
Power
Trans. Recv.
sir (1)
Frame
error (2)
arq
rrq (3)
tcp
Internet
tcp
ack (4)
Figure 4.1: System overview: A terminal (on the left), attached to a wireless link,
transmits data to a receiving node attached to the wired Internet. There are four
feedback loops aecting the delay distribution and throughput: Inner loop power
control (1), outer loop power control (2), link-layer retransmissions (3), and end-
to-end congestion control (4).
network congestion, not transmission errors, and tcp does not work well with high
packet loss rates. But achieving a small loss rate comparable to that of a wired
link is not sucient for tcp to work as well over wireless links as over the wired
Internet; performance problems have been observed also with wireless links that use
local retransmissions. As seen from the tcp end points, the dierence between such
a link and a wired one is no longer the loss rate, but the packet delay distribution.
In this chapter we derive the packet delay distribution of a power controlled
channel, the probabilities of triggering spurious timeout and spurious fast retrans-
mit in tcp, and the corresponding degradation in tcp throughput. For concrete-
ness, particular choices of various model parameters are made throughout the anal-
ysis, but it is straight-forward to apply the same analysis for other choices of pa-
rameters.
The results in this chapter seem to conrm recent simulation studies using
more detailed models for the wireless link [27]. Much of our analysis focus is on
the distribution of the end-to-end roundtrip delay, so we would like to compare our
results to empirical studies of this distribution, for both wired and wireless networks.
For the wired case, one study found that the rtt distribution can be modelled as
a shifted Gamma-distribution [34]. Empirical studies of the rtt distribution for
wireless links seem to be rare, and also for the wired Internet, the delay distribution
is not well understood.
As a motivation for our work, we believe that the engineering freedom we have
4.2. RADIO MODEL 39
tcp
ip
Link
sir
Power
tcp congestion control
Retransmission of damaged frames
Outer loop of power control
Inner loop of power control
Figure 4.2: Several layers in the networking stack interact. Feedback control is
used to reduce the inuence lower layers have on layers above, and thereby support
layer separation.
in designing the link-layer processes, such as the scheduling of retransmitted frames,
should be used to make the link friendlier to tcp. The next chapter discusses initial
work along this line. To be able to design the link-layer mechanisms in a systematic
way, good models for the link-layer processes, and their interaction with tcp, are
essential.
When using tcp over a wireless link, there are several interacting control sys-
tems stacked on top of each other, as illustrated in Figures 4.1 and 4.2. At the
lowest layer, the transmission power is controlled in order to keep the signal to
interference ratio (sir) at a desired level. This is a fast inner loop intended to
reject disturbances in the form of fading, or varying radio conditions. On top of
this, the outer power control loop tries to keep the frame error rate constant, by
adjusting the target sir of the inner loop. Next, there are local, link-layer, retrans-
missions of damaged frames. Finally, we have the end-to-end congestion control of
tcp.
4.2 Radio model
Data is transmitted over a radio link, e.g., the wcdma radio link used in umts, as
a sequence of radio frames. One radio frame corresponds to a transmission time
interval (tti) of 10 or 20 ms. Depending on the bandwidth (typically from 64 kbit/s
to 384 kbit/s) the size of a frame can vary from 160 octets (the small 10 ms tti is
not used for the lowest data rates) to 960 octets.
The transmission of the radio frames is lossy. Let p denote the overall probability
that a radio frame is damaged. The power of the radio transmitter is controlled,
so that the loss probability stays fairly constant. The target frame error rate,
p
ref
, is a deployment tradeo between channel quality and the number of required
base stations. For umts the target frame error rate is often chosen to be about
10% [13, 32]. In the following, we thus assume p
ref
= 0.1.
Power control
The typical sir-based power control uses an inner loop that tries to keep the sir
close to a reference value sir
ref
. This loop, marked (1) in Figure 4.1, often has
a sample frequency of 1500 Hz, and a one bit feedback that is subject to a delay
of two samples, i.e., 1.3 ms. The inner loop is often able to track the reference
sir
ref
within 2-3 dB, with a residual oscillation due to the delay and the severe
quantization of the power control commands in the inner loop [21]. The period of
this oscillation is typically less than 5 samples, i.e., 3.3 ms.
As there is no simple relationship between the sir and the quality of the radio
connection, there is also an outer loop that adjusts sir
ref
. The sir-updates are
controlled by the block pc in Figure 4.1. This loop uses feedback from the
decoding process; in this thesis we assume that the power control outer loop is
based on frame errors. The outer loop uses a sampling interval of one tti, one
order of magnitude slower than the inner loop. In control theory, such use of a fast
inner loop and a slower outer loop is called a cascaded control system.
As it is hard to estimate the frame error rate accurately, in particular if the
desired error rate is small, one common approach is to increase sir
ref
signicantly
when an error is detected, and decrease sir
ref
slightly for each frame that is received
successfully. It is interesting to note that this strategy resembles the tcp additive
increase, multiplicative decrease congestion control strategy. In the next section,
we discuss this approach in more detail.
Markov model
The outer loop of the power control sets the reference value for the sir. Given
a particular reference value sir
ref
= r, the obtained sir can be modelled as a
stochastic process. Together with the coding scheme for the channel, we get an
expected probability for frame errors. If the coding scheme is xed, the probability
of frame errors is given by a function f(r) which can be computed from models of
the channel and coding. Let us consider binary phase shift keying (bpsk). Then
we get the approximation
f(r) 1
_
1 Q(
_
2e
r+
2
2
/2
)
_
L
where is the coding gain, = ln 10/10, is the standard deviation of the received
sir, L is the number of bits in a radio frame, and Q(x) =
1
2
_
x
e
x
2
dx [40]. We
take = 4, also from [40], = 1 dB, corresponding to a small sir oscillation as
in [21], and N = 2560, corresponding to a modest link capacity of 128 kbit/s. This
function f(r) is shown in Figure 4.3.
In general, f(r) is a decreasing, threshold-shaped function. For small enough r,
f(r) 1, i.e., almost all frames are lost. For large enough r, f(r) 0, i.e., almost
no frames are lost. The operating point of the power control is the point close to
the end of the threshold, where f(r) = p
ref
, marked with an asterisk in Figure 4.3.
4.2. RADIO MODEL 41
0 1 2 3 4 5 6
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Prob.
sir
ref
[dB]
Figure 4.3: Frame loss probability as a function of sir

ref
. The star marks the
operating point f(r) = p
ref
.
It is the shape of f(r) close to this point that is of primary importance for the
power control behavior. Note that for control purposes only a crude approximation
of the true f(r) is necessary.
The outer loop of the power control uses discrete values for sir
ref
. One way to
keep the frame error probability close to the desired probability p
ref
is to change
sir
ref
by xed steps, based on a xed step size . Whenever a radio frame is
received successfully, sir
ref
for the next frame is decreased by . Whenever a radio
frame is damaged, sir
ref
for the next frame is increased by K. This mechanism
yields an average of one lost frame out of 1 + K, which implies p = 1/(1 + K).
We consider the case that K is a positive integer, and in particular, K = 9 which
implies p = 0.1 = p
ref
. The value of is an important control parameter, which
inuences the performance of the power control. For simplicity, we consider a xed
, even if some systems may allow the step size to be adjusted in real time.
In [40], the varying sir
ref
was modelled as a discrete Markov chain. The mean
and variance of the received sir was derived, and the dependency on the power
control step size and on the inner loop performance was investigated. We use
the same Markov model, but focus on the resulting frame error process and its
implications for tcp.
1 2 3 4 5 6 7 8 9 10 11 12
Figure 4.4: Power control Markov chain, for N = 12 and K = 9. Each state k
corresponds to a sir
ref
value of r
0
+k. There are at most two possible transitions
from state k: to k + 9 and to k 1.
To obtain a nite Markov chain for a given function f(r) and a step size , we
rst modify f by setting
f(r) =
_
1 r < r
min
0 r > r
max
Figure 4.4 illustrates the case N = 12 and K = 9. State k, 1 k N, corresponds
to r = r
0
+ k, with r
0
and N chosen such that r
0
+ < r
min
and r
0
+ (N
K + 1) > r
max
. Dene f
k
= f(r
0
+ k). Then f
k
= 1 for k 1 and f
k
= 0 for
k > N K (for convenience, we dene f
k
for all integers k, even if it is 1 k N
that is of primary interest). The transitions from state k are as follows: If k = 1,
then the next state is always 1+K. If 2 k N K, then the next state is k +K
with probability f
k
(transmission fails), and otherwise k 1 with probability 1f
k
(transmission succeeds). Finally, if N K < k N, then the next state is always
k 1.
Let
k
denote the stationary probability for state k. It is straightforward to
compute
k
, k = 1, . . . , N, as the solution to the linear equations
k
= (1 f
k+1
)
k+1
1 k K
k
= f
kK
kK
+ (1 f
k+1
)
k+1
K < k N 1
N
=
NK
f
NK
N
k=1
k
= 1
Figure 4.5 shows the stationary distribution for three values of . Here, r
0
=
r
min
= 0 dB, r
max
= 6 dB, K = 9, and N = 1 +6/. Note that a smaller step size
gives a more narrow distribution, and a smaller average sir, which means that there
is a tradeo between energy eciency and short response times. When the radio
conditions for one user change so that the user needs a higher sir, then a small
implies that the power control will adjust slowly, and in the mean time, the user will
experience a poor quality. On the other hand, a large implies a higher average
transmission power for all users, and hence more interference between users. This
interference reduces the number of users that can be served simultaneously, i.e., the
capacity of the cell.
The threshold function f(r) is also shown in Figure 4.5 (rescaled to t). For
a large , the power control state, i.e., sir
ref
, will move quite far along the tail
4.2. RADIO MODEL 43
0 1 2 3 4 5 6
0
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
0.1
f(r)
0.2dB
0.06dB
0.02dB
Prob.
sir
ref
[dB]
Figure 4.5: Stationary distribution for the power control. Each mark represents
one state of the power control, the corresponding value of sir
ref
, and its stationary
probability
i
. The dotted curve is the threshold-shaped function f(r), scaled to
t in the gure, which represents the frame error probability as a function of sir
ref
.
The star marks the desired operating point.
of the f(r) threshold. For a smaller , the control state will stay close to the
operating point where f(r) = p = 0.1. In the latter case, a linearization of f(r)
about the operating point gives a good approximation of the dynamics generated
by the Markov chain.
Radio frame loss process
Let e
t
denote a realization of the radio frame loss process: e
t
= 0 if the frame sent
in time slot t is received correctly, and e
t
= 1 if the frame is damaged. Dene
s
t
= 1 t + (K + 1)
t1
i=1
e
i
Then s
t
is the change in power control state from time 1 to time t, and it is thus
just another way to express that the state moves K steps upward when a frame is
lost, and one step downwards when a frame is received successfully.
Consider a nite segment of the loss process, e
1
, e
2
, . . . , e
, of length . If we
assume that at t = 1, the initial state of the power control is distributed according
to the stationary distribution
i
, then the probability that the realization e
1
, . . . e
occurs can be computed by conditioning on the initial state i:

P(e
1
, . . . , e
) =
N
i=1
t=1
_
e
t
f
i+s
t
+ (1 e
t
)(1 f
i+s
t
)
_
We have done an exhaustive calculation of these probabilities for all loss se-
quences of length up to 20. In the following sections, we will use these probabilities
to investigate the experience of an ip packet, or a sequence of ip packets, that
traverses the link.
4.3 Link-Layer Retransmissions
The simplest way to transmit ip packets over the wireless link is to split each ip
packet into the appropriate number of radio frames, and drop any ip packet where
any of the corresponding radio frames were damaged. But as is well-known, tcp
interprets all packet drops as network congestion, and its performance is therefore
very sensitive to non-congestion packet drops. An ip packet loss probability on the
order of 10% would be highly detrimental.
There are several approaches to recover reasonable tcp performance over wire-
less links. In this chapter we concentrate on a local and common mechanism: Auto-
matic repeat request (arq). The link receiver detects frame damage, and requests
that the damaged frame be retransmitted over the link. Use of arq is an option
in standard wireless network protocols, see [4] for an evaluation of these options in
the is-2000 and is-707 rlp standards. The eect of link-layer retransmissions is to
transform a link with constant delay and random losses into a link with random
delay and almost no losses.
There are several possible arq schemes. We will consider one of the simpler,
the (1, 1, 1, 1, 1)-negative acknowledgement scheme [27], which means that there
are ve rounds, and in each round there is a single retransmission request. When
the receiver detects that the radio frame in time slot t is damaged, it sends a
retransmission request to the sender. The frame is scheduled for retransmission
in slot t + 3 (where the delay 3 is called the rlp nak guard time). If also the
retransmission results in a damaged frame, a new retransmission request is sent
and the frame is scheduled for retransmission in slot t + 6. This goes on for a
maximum of ve retransmissions.
In the next section, we put together the frame loss process and the retransmis-
sion scheduling, to derive ip layer properties of the link. The same methods can be
applied to other retransmissions schemes than the (1, 1, 1, 1, 1)-scheme considered
here.
4.4. TCP/IP PROPERTIES 45
ip input: 1 2
Radio frames: 1 1 1 2 2 2
ip output: 1 2
d
1
d
2
Figure 4.6: Overlaying ip transmission on top of the frame loss process and frame
retransmission scheduling. Two ip packets, corresponding to n = 3 radio frames
each, are transmitted over the radio link. The second and sixth frames are damaged,
crossed out in the gure, and scheduled for retransmission three frames later. The
resulting retransmission delay for the i:th ip packet is denoted d
i
.
4.4 TCP/IP Properties
When transmitting variable size ip packets over the link, each packet is rst divided
into x size radio frames. We let n denote the number of radio frames needed for
the packet size of interest. For the links we consider, we have n 10, where n = 10
is used if we send packets of 1500 octets over a slow link with the smallest frame
size of 160 octets.
The outermost feedback loop we consider is the end-to-end congestion control
of tcp, marked (4) in Figure 4.1. The role of tcp is to adapt the sending rate to
the available bandwidth. The inputs to the tcp algorithm are measured roundtrip
time and experienced loss events, both of which depend on the links traversed by
the tcp packets. To understand and predict the behavior of tcp, we must consider
how it interacts with the power control and the link-layer retransmissions.
For any sequence of incoming ip packets, and for any nite loss realization, we
can overlay the ip packets on top of the frame sequence to see which frames are
retransmitted, and when each ip packet is completed. This process is illustrated in
Figure 4.6. Our results for properties of interest, such as the ip packet delay distri-
bution, and the probability of reordering events, are thus obtained by conditioning
on the realization of the frame loss process, and a summation over all possible loss
sequences of some nite length. Sequences of length 20 have been sucient for
our purposes, in the sense that the input sequences we have considered are short
enough that they, with high probability, can be successfully transmitted within 20
frames.
IP packet delay
One important link property for tcp is the delay distribution of packets traversing
the link. Using the loss sequence probabilities from Section 4.2, we can explicitly
compute the ip packet delay distribution. The distribution for = 0.06 dB and
increasing values of n is illustrated in Figure 4.7. The expected value and standard
deviation for three values of the step size are shown in Table 4.1. Only the delay
P
Delay [ms]
n = 1
0 60 120
+ 4
P
Delay [ms]
n = 2
0 40 60 100 120
+ 4
P
Delay [ms]
n = 3
0 20 40 60 80 100 120
+ 4
.
.
.
P
Delay [ms]
n = 6
0 20 40 60 80 100 120
+ 4
P
Delay [ms]
n = 7
0 20 40 60 80 100 120
+ 4
Figure 4.7: Retransmission delay distribution. Each bar shows one possible retrans-
mission delay, and the corresponding probability. These probabilities are calculated
for = 0.06 dB, and for packet sizes n = 1, 2, 3, 6, 7. The tti is 20 ms. For each
distribution, the value of + 4 is also shown; the tail to the right of this mark is
related to spurious tcp timeout events.
4.4. TCP/IP PROPERTIES 47
= 0.2 = 0.06 = 0.02
n = 1 0.310.94 0.320.99 0.331.03
2 0.511.08 0.531.15 0.541.20
3 0.621.08 0.641.18 0.651.24
4 0.721.09 0.741.21 0.761.28
5 0.831.09 0.851.23 0.871.31
6 0.941.09 0.961.25 0.981.35
7 1.061.09 1.071.27 1.091.38
8 1.171.09 1.181.29 1.191.41
9 1.281.09 1.291.30 1.301.44
10 1.391.09 1.401.31 1.411.46
Table 4.1: Mean and standard deviation of the ip packet delay distribution, mea-
sured in units of tti.
due to retransmissions is included (an n-frame packet is of course also subject to a
propagation delay of n tti).
The mean delay increases almost linearly with the packet size n, which is natural
since the expected number of damaged frames is proportional to n. The standard
deviation increases more slowly; one explanation is that the repairing power of a
single retransmitted frame is fairly large, independent of the packet size.
The step size seems to mostly inuence the deviation, not the mean. There
is also a signicant dependence on when looking at individual probabilities, i.e.,
the individual bars in Figure 4.7.
TCP fast retransmit
If a packet in a tcp stream is delayed so much that it lags behind three other packets
(see Figure 4.8), the acknowledgements for the next three packets are duplicated,
and the sender will go into fast retransmit/fast recovery mode. That means that
from the point of view of tcp and its congestion control, the packet is lost.
Computing the precise number of fast retransmit events is not trivial. However,
we can estimate the probability that fast retransmissions occur as follows. Consider
a randomly chosen time, when the power control state is governed by the stationary
distribution
i
. Assume that four ip packets, each of size n radio frames, arrive to
the radio link back-to-back. We then consider all possible loss sequences, and sum
the probabilities of the sequences which cause the rst packet to be delayed long
enough to be the last of the four packets to be received successfully and completely.
This gives us the probability of spurious fast retransmit, denoted P
FR
. Our results
are displayed in Table 4.2.
The probability decreases rapidly as the packet size increases. To understand
why, consider the delay d, measured in number of frames, suered by the rst packet
in Figure 4.8. A necessary condition for the packet to lag behind the next three
Sent: 1 2 3 4
Received: 2 3 4 1
Time
Figure 4.8: Reorder triggers fast retransmit.
packets, is that d 3n. For all n, the probability P(d > x) decays rapidly with x,
while both the mean and standard deviation for the delay grows only slowly with
n (see Table 4.1). Then P
FR
P(d > 3n) must also decay rapidly with n.
Therefore, spurious fast retransmissions will not be a problem for larger packet
sizes. We also see a clear dierence between small and large step sizes , this
inuence is discussed further in Section 4.5.
Timeout
A timeout event occurs when a packet, or its acknowledgement, is delayed too long.
Let rtt
k
denote the roundtrip time experienced by packet k and its corresponding
acknowledgement. The tcp algorithm estimates the mean and deviation of the
roundtrip time [24]. Let

rtt
k
and
k
denote the estimated roundtrip time and
deviation, based on measurements up to rtt
k
. tcp then computes the timeout
threshold for the next packet as
rto
k+1
=

rtt
k
+ 4
k
which means that the probability that packet k causes a timeout is given by
P
TO
= P(rtt
k
> rto
k
)
We assume that the values rtt
k
are identically and independently distributed
according to the delay distribution given in Section 4.4. For simplicity, we also
assume that the estimates

rtt
k
and
k
are perfect and equal to the true mean and
standard deviation of rtt
k
. Under these assumptions, rto
k
is constant, and we
can omit the subscript.
In tcp, timeout is the last resort recovery mechanism, and in order to get rea-
sonable performance, a timeout must be a very rare event. The value of P
TO
will
of course depend on the delay distribution. For a Gaussian or uniform distribu-
tion, P
TO
is very small (this is discussed further in the next chapter). Calculating
the P
TO
from the delay distribution of Section 4.4 results in the probabilities in
Figure 4.9. We see that the probability for spurious timeout is signicant for all
packet sizes.
The dependence on step size and packet size looks complicated, but can be
understood from Figure 4.7. The P
TO
is the tail of the distribution to the right
4.5. THROUGHPUT DEGRADATION 49
P
FR
= 0.2 = 0.06 = 0.02
n = 1 0.24% 0.63% 0.79%
2 0.00% 0.02% 0.05%
3 0.00% 0.00% 0.00%
.
.
.
.
.
.
.
.
.
.
.
.
Table 4.2: Probability of spurious fast retransmit.
of the + 4 mark in Figure 4.7. When n is increased, each of these probabilities
grows slowly, but at the same time the +4 mark moves to the right. For n 6,
the P
TO
is the sum of the probabilities for 120 ms delay and up, and each of these
grow with n. But for n = 7, +4 > 120, so the bar at 120 ms no longer contributes
to the P
TO
. And since the probability for a 120 ms delay is approximately 0.6%,
the P
TO
drops by 0.6 percentage points when n is increased from 6 to 7, as can be
seen in Figure 4.9 for = 0.06 dB.
4.5 Throughput Degradation
In this section, we derive an estimate for the performance degradation due to either
of spurious fast retransmissions and spurious timeouts. The relative degradation
depends only on the probability of the event in question, the type of event, and the
number of packets in the maximum congestion window. We illustrate the procedure
by calculating the degradation for an example link.
When computing tcp throughput, there are two distinct cases: Depending on
the bandwidth-delay product, throughput can be limited either by the bandwidth
of the path across the network, or by the maximum tcp window size.
If the product of the end-to-end roundtrip delay and the available bandwidth is
small, compared to the maximum tcp window size, spurious timeouts and spurious
fast retransmit need not lead to any performance degradation. A modest buer
before the radio link will be enough to keep the link busy even when the sender
temporarily decreases its sending rate. On the other hand, if the bandwidth-delay
product is larger than the maximum tcp window size, throughput is decreased.
The dierence between these two cases can be seen for example in the performance
evaluation [27]: In the scenarios that have a large maximum window size compared
to the bandwidthdelay product, we get a throughput that is the nominal radio link
bandwidth times 1 p (where p is the average frame loss probability), and there is
no signicant dierence between dierent link retransmission schemes. Only when
bandwidth or delay is increased, or the maximum window size is decreased, do we
see large changes in throughput when the frame error rate or retransmission-scheme
varies.
Therefore, we will concentrate on the case of a large bandwidth-delay product.
For a concrete example, consider the following scenario, illustrated in Figure 4.10:
0 2 4 6 8 10
0.0
0.2
0.4
0.6
0.8
1.0
1.2
P
TO
[%]
n
= 0.02
= 0.06
= 0.2
Figure 4.9: Probability of spurious timeout as a function of n, the number of radio
frames per ip packet, for three dierent values of the power control step size .
Radio link bandwidth 384 kbit/s, packet size m = 1500 bytes, maximum tcp
window size w = 5m = 7500 bytes, and a constant roundtrip delay time, excluding
the radio link itself, of 0.2 s. We also consider a smaller packet size of m = 960,
with w = 8m = 7680 bytes, making each ip packet t in a single radio frame. For
simplicity, we always choose w as a multiple of m.
Let T denote the average end-to-end roundtrip time, 0.21 s for the larger
packets and slightly less (0.206 s) for the smaller packets. This implies that the
bandwidth-delay product is larger than the maximum window size, with an ideal
throughput w/T of 34.8 Kbyte/s for the case of large packets and somewhat larger,
36.3 Kbyte/s, for the case of small packets. Note that both values are smaller than
the available radio bandwidth of the link, 384000(1 p
ref
)/8 42.2 Kbyte/s.
Fast retransmit followed by fast recovery, and timeout followed by slow start,
can be treated in a uniform way. Assume that fast retransmit (or timeout) oc-
curs independently with a small probability P
LOSS
. Consider a typical cycle start-
ing with the event that causes the performance degradation, followed by 1/ P
LOSS
packets that are sent without timeouts or retransmissions. We want to compare the
throughput during this cycle with the ideal throughput w/T. For simplicity, use
only congestion windows that are an integral number of packets, and also assume
that the cycle length N = 1/ P
LOSS
is an integer. The cycle can be divided into
two phases: An initial recovery phase of j roundtrip times, when packets are sent,
followed by the second phase of (N )m/w roundtrip times at full speed, w/m
4.5. THROUGHPUT DEGRADATION 51
pc
sir
ref
+
Power
Trans. Recv.
sir (1)
Frame
error (2)
arq
rrq (3)
tcp
Internet
tcp
ack (4)
384 kbit/s
Delay = 0.2 s
mtu = 1500 bytes
max cwnd = 7500 bytes
Figure 4.10: Numerical example.
packets each roundtrip time, N packets in all.
The throughput during one cycle can be expressed as
throughput =
N
j + (N )m/w
m
T
Compared to the ideal throughput w/T, the relative degradation is
_
w
T
throughput
_
_
w
T
=
d
d + N
=
1
1 + 1/(d P
LOSS
)
where we have substituted d = jw/m , the number of additional packets that
would have been sent during the cycle if tcp had never decreased the congestion
window.
Degradation due to fast retransmit
First consider the eect of spurious fast retransmit and fast recovery. Here, the
sending tcp will resend the delayed packet (thinking that it is lost), and halve its
congestion window. The window size is then increased by m each rtt, until it
reaches the maximum value, w.
For m = 1500, w = 7500 = 5m, the congestion window is halved to w/2 and, to
keep the window size as an integral number of packets, we round this down to 2m.
During the next three roundtrip times tcp sends 2, 3, 4 packets, or 9 packets total,
after which it is up to full speed again, sending 5 packets each roundtrip time. This
means that = 9, j = 3 and d = 6: Due to the spurious fast retransmit event,
Degradation Packet size = 0.2 0.06 0.02
due to P
FR
m = 960 2.3% 5.9% 7.3%
1500 0.0% 0.1% 0.3%
due to P
TO
m = 960 5.6% 13.0% 16.5%
1500 4.9% 6.4% 8.4%
Table 4.3: Performance degradation, due to spurious fast retransmit and spurious
timeout.
6 packets less are sent during the cycle. The resulting performance degradation
is 1/(1 + 1/(6 P
FR
)). As each packet corresponds to two radio frames, n = 2 in
Table 4.2. The performance degradation is noticeable only for the smaller step sizes.
With = 0.06, P
FR
0.02% yields N = 5000 and a performance degradation of
6/5006 0.1%, while = 0.02, P
FR
0.05 yields N = 2000 and a performance
degradation of 6/2006 0.3%.
For the packet size m = 960, the situation is dierent. The maximum window
size is w = 8m, which is halved at spurious fast retransmit. During the 4 roundtrip
times of the recovery phase, tcp sends 4, 5, 6 and 7 packets, giving d = 10.
The large step size, = 0.2 yields P
FR
0.24%, N = 417, and a performance
degradation of 10/427 2.3%. For the smaller step sizes = 0.06 and 0.02, the
performance degradation is worse, 5.9% and 7.3%, respectively.
These gures are summarized in Table 4.3. The degradation is signicant for
the small packet size, and much less of a problem for larger packets.
Degradation due to timeouts
Next, consider the eect of timeout. When recovering from timeout, the slow start
threshold is set to w/2. The congestion window is reset to one packet, and doubled
each roundtrip time, until it reaches the slow start threshold. Above the slow start
threshold, the usual increment of the congestion window of one packet per roundtrip
time is used until we get back to the maximum value w.
For m = 1500, w = 5m = 7500, the length of the recovery phase is 4 roundtrip
times, during which we send 1, 2, 3 and 4 packets. We get d = 10, and from the
P
TO
calculated in Section 4.4 we get a performance degradation of 4.9%, 6.4%,
and 8.4% for the three step sizes.
For m = 960, w = 8m = 7680, the length of the recovery phase is 6 roundtrip
times, during which we send 1, 2, 4, 5, 6, 7 packets. Thus, d = 23, and we get a
performance degradation of 5.6%, 13.0% and 16.5% for the three step sizes.
These gures are also shown in Table 4.3. We see that degradation due to
spurious timeout is signicant for all considered values of the packet size and power
control step size.
4.6. SUMMARY 53
Inuence of the power control step size
The choice for the step size is determined by the tuning of the power control. It
is is a tradeo between short response times and energy eciency, as discussed in
Section 4.2. However, as can be seen in Table 4.3, the choice of also inuences
tcp performance. For both spurious timeout and spurious fast retransmit, we see
that the degradation gets worse as the power control step size is made smaller.
To understand why, note that a small step size implies a low correlation in the
frame loss process (in the limit 0, frame errors become independent). Then
we have almost the same loss probability for original transmissions and for frame
retransmissions. On the other hand, a large step size implies a signicant increase in
transmission power after each frame loss, which means that each retransmission of
a frame will use a signicantly higher sir than the previous attempt. This reduces
the number of retransmissions, which in turn reduces the retransmission delays and
the problems with spurious fast retransmit and spurious timeout.
4.6 Summary
The properties of the lossy radio link considered in this chapter reduces tcp
throughput. By modeling the underlying processes controlling transmission power
and retransmission scheduling on the link, we can derive the link properties, in par-
ticular the delay distribution, that are relevant for tcp. We found that when using
ip packets that t in one or two radio frames, as is common for high-bandwidth
wireless links, the link will trigger spurious fast retransmissions in tcp. Spuri-
ous timeouts will also occur, and this eect is signicant for both large and small
packets.
Chapter 5
Improving the link layer
This chapter proposes a measure of tcp-friendliness, and we see how
adding carefully chosen additional delays can make a link work better
for tcp.
5.1 Introduction
We believe that, as far as possible, the link-layer should be engineered to be tcp-
friendly, reducing the dierences between wired and wireless links. This chapter
investigates ways to optimize the operation of link-layer processes, in order to im-
prove tcp performance, and it consists of two mostly independent parts.
Section 5.2 proposes a model for analyzing and optimizing the retransmission
scheduling. It is intuitively appealing to view retransmissions as a form of feedback,
but non-trivial to nd the right way to express this idea in a mathematical form.
The main part is Sections 5.35.5. These sections discuss modications to the
wireless link of the previous chapter, to reduce the tcp performance degradation
due to spurious timeouts. The timeout procedure in tcp motivates a denition of
a quality measure of tcp friendliness. The main result is that tcp performance
can be improved by adding carefully chosen articial delays to certain packets.
There will naturally be some residual idiosyncrasies of wireless channels that
cannot be dealt with in the link-layer; the work presented in this chapter should be
viewed as complementing both developments to make tcp more robust to strange
links, and cross-layer developments that let the link and the end-nodes exchange
information about link and ow properties.
5.2 Retransmission as feedback
To be able to analyze the impact of the scheduling mechanism on link properties
such as the delay distribution, it is of interest to model the retransmission scheme.
Feedback is an intrinsic property of the retransmission mechanism. Below we pro-
55
56 CHAPTER 5. IMPROVING THE LINK LAYER
Delay
Scheduling
Delay
x(k) t(k)
Channel Sorting
y(k)
e(k)
Figure 5.1: Retransmission model.
pose a model for the relationship between the input frames, the frame error process,
and the in-order output frames, where this feedback is explicitly shown. We believe
that this model will be useful for further studies of retransmission scheduling.
Let k denote time in units of the transmission time interval (tti), and consider
the following input and output signals, also shown in Figure 5.1.
x(k) = # of input frames up to time k
e(k) = # of errors up to time k
y(k) = # of in-order output frames up to time k
These are accumulated rate functions, also used in network calculus, and increasing.
Also dene
t(k) = Index of frame transmitted at time k
The function t(k) is not increasing.
Consider a simple one-parameter family of retransmission schemes, where each
damaged frame is retransmitted g slots later, and there is no limit on the number of
times a frame may be resent. The parameter g corresponds to the rlp nak guard
time.
To describe the process mathematically, start with the queue at the input to the
scheduler. Let s(k) be the number of time slots up to time k that are not used for
retransmissions, and let f(k) be the number of frames that have been transmitted
(but not necessarily received successfully) up to time k. Then
s(k) = k e(k g)
f(k) = min
k
(x() + s(k) s())
where the minimum in the latter equation is obtained when is the start of the
current busy period.
The scheduling can be described as
t(k) =
_
f(k) if e(k g) = e(k g 1)
t(k g) if e(k g) > e(k g 1)
(5.1)
5.3. IP PACKET DELAY 57
Delay [ms] 0 40 60 100 120 160 180
Probability [%] 80.6 8.8 9.3 0.6 0.6 0.03 0.03
Table 5.1: Delay distribution of an example wireless link.
The rst case corresponds to an original transmission, and the second case to a
retransmission. Finally, y(k) is dened by y(k) = n if all frames up to n have been
received properly at time k, but frame n + 1 has not. In symbols,
y(k) = max{n : n, j k, t(j) = , e(j) = e(j 1)}
So where is the feedback? It is included explicitly, in (5.1). This model, together
with a model for the stochastic process e(k), lets us optimize the parameterized
retransmission scheme. The delay at time k is dened by
d(k) = min{ 0 : y(k + ) x(k)}
If x and e are stationary processes, with average rates that sum to less than one,
then also d(k) is a stationary process, and its properties can, at least in principle,
be calculated from x, e, and the retransmission model. If Q(d) is a quality measure
that depends on the properties of d, one can formulate the optimization problem
g
= arg max
g
Q(d)
which gives the optimal value for the retransmission delay.
Intuitively, we expect that g
will depend on the autocorrelation of e; it seems

reasonable to use a retransmission delay such that the correlation, between loss of
the original transmission and loss of the retransmission, is small.
This one parameter retransmission model is quite limited. Other schemes can
be modeled analogously, as long as the relation between e and s is simple, and the
scheme does not need an additional queue for retransmitted packets. The challenge
is to nd a powerful but simple parameterization of an interesting class of schemes.
We now leave the subject of retransmission scheduling, and develop a particular
quality measure.
5.3 IP packet delay
Since a link employing link-layer retransmission yields a very small packet loss
probability, the most important characteristic of the link is the packet delay distri-
bution. If the distribution is suciently friendly to tcp, then the layering of the
system works nicely, which means that upper layers like tcp need not be aware of
any particular properties of individual links in the network.
In the previous chapter, we computed the packet delay distribution explicitly,
for a channel with sir-based power control. The resulting delay distribution for our
example channel, with a power control step size = 0.06 dB and n = 2 is shown in
Table 5.1. This table includes only the delays due to radio frame retransmissions,
there is also a xed delay of 40 ms for the transmission of two radio frames.
TCP performance degradation
In observations and performance evaluations of tcp over wireless links [27], the
properties of a wireless link can shine through to the tcp layer in three dierent
ways:
Genuine packet loss: With bad enough radio conditions, packet drops are in-
evitable. We will not consider such genuine packet loss here, as we assume
that the radio channel is good enough that the power control and link-layer
transmissions can get packets through.
Packet reorder: For a link with highly variable delay, packets can get reordered.
Severe reordering can trigger a spurious tcp fast retransmit.
Spurious timeout: A packet that is not lost, only severely delayed, can trigger a
spurious tcp timeout.
Spurious timeout
A tcp timeout event occurs when a packet, or its acknowledgment, is delayed too
long, as determined by the timeout value rto. rto is computed from estimates of
the mean and variance of the rtt, as described in the previous chapter.
An idealized model of tcp is to assume that the estimation is perfect, so that
the timeout value is set as
rto = (rtt) + 4(rtt)
where () and () denote the mean and standard deviation. Then the spurious
timeout probability is given by
P
TO
(rtt) = P(rtt > (rtt) + 4(rtt)) (5.2)
Note that P
TO
is invariant under addition of constant delays.
For the delay distribution of Table 5.1, we get rto 103 ms and the probability
that the delay is larger is P
TO
0.68%. In the previous chapter, we had a spurious
timeout probability on the order of 0.5%1%, when varying the packet size n and
power control step size , and we saw that even a fairly small P
TO
can result in
signicant performance degradation.
5.4 Improving the link-layer
It is not trivial to dene precisely what properties a link should have in order to
be friendly to tcp. It seems clear that for example links with normal or uniformly
5.4. IMPROVING THE LINK-LAYER 59
distributed and independent delays are friendly enough. We dene a measure of
tcp-friendliness by applying (5.2) to an arbitrary stochastic variable X, repre-
senting the independent and identically distributed packet delays. We also dene
rto(X) as the corresponding timeout value.
rto(X) = (X) + 4(X)
P
TO
(X) = P(X > rto(X))
Let us compare the P
TO
for the wireless delay of Table 5.1, 0.68%, to the P
TO
of some other distributions:
If X is uniformly distributed on an interval a X b, then it has mean
= (a + b)/2, and standard deviation = (b a)/(2
3), so that + 4 =
b + (2/
3 1/2)(b a) > b. Hence, P

TO
(X) = 0.
If X is Gaussian, then timeouts will be rare: P
TO
(X) =
1
2
_
4
e
x
2
/2
dx
0.003%.
For an arbitrary distribution with nite mean and variance, Chebyshevs in-
equality, P(|X | a) 1/a
2
, yields the bound P
TO
6.25%.
We see that the rst two distributions, which we know are friendly to tcp, yield
a P
TO
at least two orders of magnitude below the worst case given by Chebyshev.
The wireless delay yields a signicantly higher P
TO
, although still with some margin
to the worst case.
The motivation for using P
TO
as a quality measure is the calculation of the
timeout value in tcp. Timeout is intended to be the last resort recovery mechanism,
and for tcp to work properly, spurious timeout must be a rare event. We therefore
dene a tcp-friendly link to mean a link with no loss or reorder, and with a delay
distribution that yields a small P
TO
.
It would be nice to use the P
TO
measure to optimize the retransmission schedul-
ing, as described in Section 5.2. But in the remainder of this chapter, we will follow
a much simpler route, which never the less allows us to decrease the P
TO
consider-
ably: Addition of articial delays.
Introducing additional delays
Assume that we have a discrete delay distribution X, P(X = d
i
) = p
i
, where
d
i
< d
i+1
. It is typical, but not required, that also p
i
p
i+1
.
We consider the following class of tweaks to X. For each packet that experiences
a delay X = d
i
, buer the packet so that it gets an additional delay
i
. This denes
a new distribution

X, P(
X = d
i
+
i
) = p
i
(or if it happens that d
i
+
i
= d
j
+
j
for some i = j, the corresponding probabilities are added up). For an example of
what X and

X can look like, see Figures 5.2.
What is the best choice for
i
0? One possible answer is given by the opti-
mization problem
min
i
0
(
X)
P
TO
(
X)
where > 0 is a maximum allowed value for P
TO
(
X). This means that we want to

push down our measure of tcp-unfriendliness, while at the same time not adding
more delay than necessary. We will see that after some simplications, this is a
quadratic optimization problem.
First, require that P
TO
(
X) corresponds to a tail of the original distribution X.

Let k be the smallest value such that
ik+2
p
i
. Let c = d
k+1
+ < d
k+2
, where
0 is a robustness margin. We impose the additional constraints d
i
+
i
d
k+1
for i k,
i
= 0 for i > k, and rto(
X) = c. Then, for any

i
satisfying these
new constraints, we will have P
TO
(
X) =
ik+2
p
i
. We get the optimization
problem
min
1
,...,
k
(
X)
(
X) + 4(
X) = c
0
i
d
k+1
d
i
, for i k
To write it in matrix form, let denote the vector (
1
, . . . ,
k
)
T
, and similarly for
p and d. Let S = 16 diag p 17pp
T
, b
i
= 2p
i
(16d
i
+ c 17), m
i
= d
k+1
d
i
and
= 16
2
(c )
2
, where and denote the mean and standard deviation of the
original delay X. We can then rewrite the problem as
min
1
,...,
k
p
T
T
S + b
T
+ = 0
0
i
m
i
Remarks
Since the symmetric matrix S is typically indenite, the problem is not convex.
But it can be solved in exponential time O(k
3
3
k
), which has not been a problem
thanks to the very limited size of k.
The role of can be seen in after graph in Figure 5.2; it is the margin, on
the delay axis, between the rto = + 4 and the delay value closest to the left
(120 ms).
The typical solution is of the form x = (0, . . . , 0,
j
, m
j+1
, . . . , m
k
)
T
. When the
optimum has this form, it means that the cheapest way to increase the rto, in terms
of mean delay, is to increase the
i
corresponding to the smallest p
i
. Necessary and
sucient conditions for the optimum to be of this form has not yet been determined.
5.4. IMPROVING THE LINK-LAYER 61
P
Delay [ms]
Before
0 40 60 100 120
+ 4
P
Delay [ms]
After
0 40 86 120
+ 4
Figure 5.2: Delay distribution, before and after optimization.
We aim at decreasing the distribution tail as measured by P
TO
, and pay a small
price in mean delay. This is a dierent objective than an ordinary jitter buer,
which aims for small variance, and pays with a larger mean delay.
Numerical example
Now consider the delays d
i
and probabilities p
i
in Table 5.1, and assume that packets
are independently delayed according to the given probabilities. This distribution is
also shown at the top of Figure 5.2. Before tweaking the delays, we have E(X)
10.6 ms, and P
TO
(X) 0.68 %.
With = 0.1% and = 10 ms, the above optimization procedure yields k = 4
and the optimal additional delay (0, 0, 26, 20)
T
ms. This modied distribution
is shown at the bottom of Figure 5.2. The mean additional delay is only 2.54 ms,
which seems to be a small cost, if we compare it to the propagation delay for the
packet, which is 40 ms, or the end-to-end delay which necessarily is even larger.
We also achieve P
TO
< , if fact, we actually get P
TO
0.06%.
Consider the same example scenario as in the previous chapter, illustrated in
Figure 4.10. The available radio bandwidth, 384 kbit/s with 10% losses, corresponds
to 42.2 Kbyte/s. The ideal tcp performance, one maximum window per rtt, was
34.8 Kbyte/s, which was reduced by 6.4% to 32.6 Kbyte/s due to spurious timeout
events.
For the tweaked link, we have a slightly larger rtt, which in itself would decrease
the throughput by 1%, and a signicantly smaller P
TO
. The resulting throughput
is 34.2 Kbyte/s, an improvement by 5% compared to the unmodied link, and only
1% below the ideal tcp throughput. The degradation due to spurious timeout is
almost eliminated. These gures are summarized in Table 5.2.
The important point is that a simple but carefully selected modication to the
link-layer yields a modest but signicant performance improvement.
Kbyte/s
Available radio bandwidth 42.2
Ideal tcp throughput 34.8
With wireless link 32.6
With optimized wireless link 34.2
Table 5.2: The tcp throughput for the wireless link, before and after optimization.
5.5 Robustness
In the analysis, we have so far only considered the delay variation from the wireless
link, and assumed that the delay over the rest of the network path is constant. Now
assume that the articial delays are chosen optimally based on the delay distribution
for the isolated link, and applied to the end-to-end rtt over the complete network
path. In this section we will prove that the resulting P
TO
is still small, under some
conditions on the delay distribution for the rest of the network.
If the delay in the rest of the network is modelled as a stochastic variable V , then
the end-to-end rtt is

X +V . So our objective is to nd a bound on P
TO
(
X +V ).
Since P
TO
is invariant under the addition of constant delays, we can make the
convenient assumption that E(V ) = 0.
If we assume that

X and V are independent, we can derive a simple bound
for P
TO
(
X + V ), which will also depend on the robustness margin . First, dene

c
= E(
X + V ) + 4(
X + V ). Note that c
c, so that c
x
i
d
i
for i k.
Next, condition on

X,
P
TO
(
X + V ) = P(
X + V > c
)
=
i
p
i
P(V > c
x
i
d
i
)
i=1
p
i
P(V > ) +
i>k
p
i
P(V > ) +
This bound shows that if the delay variations in the rest of the network are small
enough relative to our robustness parameter , P
TO
(
X+V ) will not be much larger

than . For typical distributions, the bound for the rst sum is very conservative.
This is because the rst few p
i
dominates, while the corresponding c
x
i
d
i
are
signicantly larger than . A more precise bound can be calculated using additional
information about p
i
and the distribution of V .
5.6. SUMMARY 63
5.6 Summary
The rst part of this chapter discusses optimization of the retransmission schedul-
ing. An input/output model is suggested where the role of the scheduling mecha-
nism is made explicit as a feedback.
The main contribution has been to show that a slight articial increase of the
delays of certain retransmitted packets may reduce the risk of spurious timeout
in tcp and hence increase the throughput; in an example the increase was 5%.
The articial delay distribution is optimized o-line and applied on-line. The ad-
ditional delay that is applied to a packet depends only on the retransmission delay
experienced by that same packet, and this information is available locally at the
link.
Chapter 6
Conclusions and further work
This chapter summarizes the results of the preceding three chapters, and
outlines directions for future research.
6.1 Radio network feedback
Chapter 3 studies tcp performance over the high-speed downlink packet access
(hsdpa) channel in the universal mobile telecommunications system (umts), i.e.,
downlink performance. The channel disturbances in our model and simulations are
bandwidth changes and temporary outages. A split-connection scheme is proposed,
analyzed, and compared to a setup with standard tcp Reno end-to-end.
The proposed scheme uses an http proxy, and explicit signalling from the ra-
dio network controller (rnc) to the proxy. The proxy does not use tcp Reno to
determine its window size and sending rate, instead it uses a controller based on
feedforward of the actual bandwidth of the radio channel, and feedback from the
queue length at the rnc. The information needed by the controller is provided by
the rnc in the form of radio network feedback (rnf) messages.
With this proxy, the users response time is shortened, and the utilization of
the radio channel is improved, compared to tcp Reno. Thanks to the feedforward
control, the proxy can adapt faster to bandwidth changes. In particular, it recovers
better from temporary outages. The proxy gets accurate rnf messages at the start
and end of an outage. In contrast, tcp Reno, like any end-to-end protocol, will see
no acks at all during an outage, and it is required to use exponential backo when
probing the channel to see when the outage ends.
The model based design allows tuning of parameters such as the target rnc
queue size q
ref
, and the amount of buering space for the queue. The proposed con-
troller uses well known principles: Feedforward, to adjust to bandwidth changes,
and feedback, to control the rnc queue size in spite of uncertainties in the avail-
able bandwidth and delay information. The actual control laws are very simple;
more advanced controllers could be designed to improve performance, or to handle
65
66 CHAPTER 6. CONCLUSIONS AND FURTHER WORK
additional disturbances in the system. One interesting extension is to introduce an
outer control loop that selects a suitable value for q
ref
.
The stability analysis in Chapter 3 is limited, it would be desirable to improve
the analysis to prove global stability of the non-linear system, including the prop-
agation delays.
Implementation of the proxy and the rnf-enabled rnc is needed, to evaluate
the performance of the proxy solution for the real system, and to compare it to
other alternatives, e.g., performance enhancing proxies.
6.2 Delay variations of wireless links
Since modern wireless links include the capability of local retransmissions of dam-
aged frames, the primary dierence between such a link and a wired one is no longer
the loss rate, but the packet delay distribution.
Chapter 4 develops a model for the frame loss process of a wireless link, using
models for the power control and retransmission processes. This model lets us derive
the link properties that are relevant to tcp, in particular the delay distribution.
To quantify the impact on tcp, the performance degradation due to spurious
timeout and spurious fast transmit was computed, and found to be up to 15%.
A simple way to alleviate the performance degradation caused by spurious fast
retransmissions is to congure the link to not forward packets out of order. It
seems clear that for links where tcp performance is important, and an option of
in-order delivery is available, it should be enabled.
An accurate model for the link-layer processes, and the resulting delay distribu-
tion, makes it possible to tune system parameters in a systematic way. The model
can also be used for the design of new controllers. A rst step in this direction
is described in Chapter 5, which addresses the problem of spurious timeout. We
dene a measure of tcp-friendliness for a link, or more precisely, for the links de-
lay distribution. The timeout value in tcp makes timeout a very rare event if the
roundtrip time has a uniform or normal distribution, but that is not the case for
the delay due to link-layer retransmissions.
A wireless link can be made more friendly to tcp, in this sense, by adding care-
fully chosen articial delays to certain packets. Choosing these additional delays,
and at the same time keep the average delay increase small, is posed as an opti-
mization problem. If the link adds the optimal delay to each packet, the average
delay increase is only a few ms, but still most of the spurious timeouts, and the
corresponding performance degradation of tcp, is eliminated.
One interesting mathematical problem is extending the delay optimization to
a continuous setting. One variation is as follows: Consider an input which is a
stochastic process x(t), a non-negative control signal d(t), and the output x(t) =
x(t) + d(t). Choose an optimal d(t), depending on the history of x(t), subject to
an objective and a set of constraints, which are all expressed as functionals of x(t).
The used models need further experimental validation. The results are consis-
6.3. APPROACHES TO TCP OVER WIRELESS 67
Internet
End-to-end
Split connection Link-layer
Figure 6.1: Approaches to the tcp over wireless problem.
tent with experiments and simulations in the literature, which focus on high level
properties such as tcp throughput. Data sets with realizations of the underlying
frame loss process would be highly useful for validation.
6.3 Approaches to TCP over wireless
As described in the introduction and illustrated in Figure 6.1, there are several
possible approaches to improved tcp performance over radio links, classied as:
End-to-end: Improve the transport algorithms in the endpoints, i.e., tcp, to
adapt better to the disturbances from wireless links.
Split-connection: Use a proxy to hide the wireless link from peers on the Internet.
Link-layer: Design link-layer tradeos that results in link properties that are eas-
ier to handle for the endpoints.
Improving the end-to-end tcp algorithms is an area of intense research. A
drawback of this approach is that deployment of new algorithms aect all Internet
end systems, which is a slow and costly process. Tuning the link properties is more
practical from a deployment point of view, at least if the tuning can be done before
widespread adoption of a new link type. For example, spurious fast retransmissions
can be eliminated by having the wireless receiver buer packets, making sure that
packets are never forwarded out of order (this in-order option is already available
on real systems [13, 32]). To reduce the number of spurious timeouts, the link can
add articial delays to packets or acknowledgements, to make the delay variation
more friendly to tcp, either random delays as in [28], or as described in Chapter 5.
Also the link-layer retransmissions themselves can be seen as an example of changing
the link properties to make the link more tcp friendly.
To this list, one could also add overall delay reduction. Smaller delays im-
proves network performance in general and tcp performance in particular. If the
bandwidth-delay product can be made smaller than the maximum window size in
68 CHAPTER 6. CONCLUSIONS AND FURTHER WORK
tcp
ip
Link
sir
Power
tcp congestion control
Retransmission of damaged frames
Outer loop of power control
Inner loop of power control
Figure 6.2: The layers in the networking stack, and the corresponding control loops.
tcp, then a modest size buer just before the radio link should be sucient to
keep the link saturated, regardless of uctuations in the tcp congestion window.
It is hard to say if reducing bandwidth-delay product is a realistic objective, as it
depends on the speed and delay of future link technologies. Extensions to tcp that
increase the maximum window size would also help, for the same reasons [25].
6.4 Towards a systematic design for layer separation
If we look at the networking stack in Figure 6.2, the ideal is to have layer separation:
Each layer should operate independently and not interact with the inner workings
of other layers. In control theory, a similar notion is cascaded control, where each
loop operates on its own time-scale, each inner loop working on a faster time-scale
than the outer loops. In each layer, there are tradeos that are inuenced by the
design of the control loops. If we want to improve layer separation by changing the
automatic control in the various layers, where should we put the eort?
In the lowest layer, the power control, the control design has severe constraints
of its own, relating to energy eciency, stability and cost of deployment. We can
probably not add new requirements, related to tcp-friendliness, to power control
design.
At the opposite end of the networking stack, end-to-end congestion control,
improvements to the tcp algorithms in the end-nodes are important, but dicult
for several reasons. Since tcp has to be general and work over any path across the
heterogeneous Internet, with many dierent types of links, it does not make sense to
tailor tcp for the peculiarities of one link type. End nodes have limited information
about what goes on in the link (note that the link and the tcp implementations
are not only in separate layers, they are also geographically separate). There are
also severe design constraints, such as fairness and global stability of the Internet.
However, we do have a fair amount of engineering freedom in the link-layer, even
if we do not want to modify the power control. In this thesis we have considered
only one out of many possible schemes for link-layer retransmissions (arq). There
are also other link-layer mechanisms that can be exploited, e.g., forward error cor-
rection (fec), rate adaptation and scheduling. Joint optimization of fec, arq and
transmission power is investigated in [6], and extended further, to also include rate
6.4. TOWARDS A SYSTEMATIC DESIGN FOR LAYER SEPARATION 69
adaptation, in [38].
How to get the most out of these dierent mechanisms, used together, is not well
understood. The output signals we want to control are the capacity, loss and delay.
It is clear that the available control signals enables us to make tradeos between
how radio disturbances aect the output signals. But how should the disturbance
rejection work be divided between power control, rate adaptation, fec, arq, and
scheduling? What are the fundamental limitations on what can be achieved?
To achieve a general layer separation between tcp and radio links, we need a
class of friendly link characteristics, in terms of the dynamics of capacity, losses
and delay, such that:
Link-layer control can be used to turn real physical radio channels into links
of this class, with close to full utilization of the radio resources.
Congestion control of end-to-end tcp can be designed to work well over all
links in the class, and over all network paths made up of links from this class.
Searching for such a class, and designing the corresponding link-layer and end-to-
end controllers, is a meet-in-the-middle attack on the tcp-over-wireless problem.
Bibliography
[1] A. A. Abouzeid, S. Roy, and M. Azizoglu. Stochastic modeling of TCP over
lossy links. In INFOCOM (3), volume 3, pages 17241733, 2000.
[2] M. Allman, V. Paxson, and W. Stevens. TCP congestion control. RFC 2581,
April 1999.
[3] G. Appenzeller, I. Keslassy, and N. McKeown. Sizing router buers. In SIG-
COMM, Portland, September 2004. ASM.
[4] Y. Bai, P. Zhu, A. Rudrapatna, and A. T. Ogielski. Performance of TCP/IP
over IS-2000 based CDMA radio links. In IEEE VTC2000-Fall. IEEE, 2000.
[5] C. Barakat and E. Altman. Bandwidth tradeo between TCP and link-level
FEC. Comput. Networks, 39(5):133150, 2002.
[6] D. Barman, I. Matta, E. Altman, and R. El Azouzi. TCP optimization through
FEQ, ARQ and transmission power tradeo. In International Conference on
Wired/Wireless Internet Communications WWIC, february 2004.
[7] J. Border, M. Kojo, J. Griner, G. Montenegro, and Z. Shelby. Performance
enhancing proxies intended to mitigate link-related degradations. RFC 3135,
June 2001.
[8] A. Canton and T. Chahed. End-to-end reliability in UMTS: TCP over ARQ.
In IEEE Globecom, 2001.
[9] G. Carneiro, J. Ruela, and M. Ricardo. Cross-layer design in 4G wireless
terminals. IEEE Wireless Communications, 11(2):713, April 2004.
[10] S. Cen, P. C. Cosman, and G. M. Voelker. End-to-end dierentiation of con-
gestion and wireless losses. IEEE/ACM Trans. on Networking, 11(5):703717,
2003.
[11] T. Chahed, A-F. Canton, and S-E. Elayoubi. End-to-end TCP performance in
W-CDMA/UMTS. In ICC, Anchorage, May 2003.
[12] A. Chockalingam, A. Zorzi, and V. Tralli. Wireless TCP performance with
link layer FEC/ARQ. In IEEE ICC, pages 12121216, June 1999.
71
72 BIBLIOGRAPHY
[13] A. Dahlen and P. Ernstrom. TCP over UMTS. In Radiovetenskap och Kom-
munikation 02, RVK, 2002.
[14] A. DeSimone, M. Chuah, and O. Yue. Throughput performance of transport-
layer protocols over wireless LANs. In IEEE Globecom93, pages 542549,
1993.
[15] J. Postel (ed.). Transmission control protocol. RFC 793, September 1981.
DARPA Internet Program.
[16] R. Braden (ed.). Requirements for internet hosts communication layers.
RFC 1122, October 1989.
[17] G. Fairhurst and L. Wood. Advice to link designers on link automatic repeat
request (ARQ). RFC 3366, August 2002.
[18] S. Floyd and T. Henderson. The NewReno modication to TCPs fast recovery
algorithm. RFC 2582, April 1999.
[19] S. Floyd, J. Mahdavi, M. Mathis, and M. Podolsky. An extension to the
selective acknowledgement (SACK) option for TCP. RFC 2883, July 2000.
[20] C. P. Fu and S. C. Liew. TCP veno: TCP enhancement for transmission over
wireless access networks. IEEE Journal on Selected Areas in Communications,
21(2):216228, 2003.
[21] F. Gunnarsson and F. Gustafsson. Power control in wireless communications
networks from a control theory perspective. In IFAC World Congress,
Barcelona, 2002.
[22] E. Hossain, D. I. Kim, and V. K. Bhargava. Analysis of TCP performance
under joint rate and power adaptation in cellular WCDMA networks. IEEE
Trans. on Wireless Comm., 3(3), May 2004.
[23] B. A. Huberman, P. L. T. Pirolli, J. E. Pitkow, and R. M. Lukose. Strong
regularities in World Wide Web surng. Science, 280(5360):9597, 1998.
[24] V. Jacobson. Congestion avoidance and control. ACM Computer Communi-
cation Review, 18:314329, 1988.
[25] V. Jacobson, R. Braden, and D. Borman. TCP extensions for high perfor-
mance. RFC 1323, May 1992.
[26] V. Kawadia and P. R. Kumar. A cautionary perspective on cross-layer design.
IEEE Wireless Communications, 12(1):311, Februari 2005.
[27] F. Khan, S. Kumar, K. Medepalli, and S. Nanda. TCP performance over
CDMA2000 RLP. In IEEE VTC2000-Spring, pages 4145, 2000.
73
[28] T. E. Klein, K. K. Leung, R. Parkinson, and L. G. Samuel. Avoiding TCP
timeouts in wireless networks by delay injection. In IEEE Globecom, 2004.
[29] R. Ludwig and R. H. Katz. The Eifel algorithm: Making TCP robust against
spurious retransmissions. ACM Computer Communication Review, 30(1), Jan-
uary 2000.
[30] S. Mascolo, C. Casetti, M. Gerla, M. Y. Sanadidi, and R. Wang. TCP West-
wood: bandwidth estimation for enhanced transport over wireless links. In
MobiCom, Rome, Italy, 2001.
[31] M. Mathis, J. Mahdavi, S. Floyd, and A. Romanow. TCP selective acknowl-
edgment options. RFC 2018, October 1996.
[32] M. Meyer, J. Sachs, and M. Holzke. Performance evaluation of a TCP proxy
in WCDMA networks. IEEE Wireless Communications, 10(5):7079, October
2003.
[33] I. Cabrera Molero. Radio network feedback to improve TCP utilization over
wireless links. Masters thesis, KTH, 2005.
[34] A. Mukherjee. On the dynamics and signicance of low frequency components
of internet load. Technical report, Univerity of Pennsylvania, 1992.
[35] C. Hava Muntean, J. McManis, and J. Murphy. The inuence of Web page
images on the performance of Web servers. Lecture Notes in Computer Science,
2093, 2001.
[36] J. Pan, J.W. Mark, and X. Shen. TCP performance and behaviors with lo-
cal retransmissions. Journal of Super Computing (special issue on Wireless
Internet: Protocols and Applications), 23(3):225244, 2002.
[37] K. Ramakrishnan, S. Floyd, and D. Black. The addition of explicit congestion
notication (ECN) to IP. RFC 3168, September 2001.
[38] C. Rinaldi. Link-layer error recovery techniques to improve TCP performance
over wireless links. Masters thesis, KTH, 2005.
[39] N. K. G. Samaraweera. Non-congestion packet loss detection for TCP error
recovery using wireless links. IEEE Proceedings-Communications, 146(4):222
230, 1999.
[40] A. Sampath, P. S. Kumar, and J. M. Holtzman. On setting reverse link target
SIR in a CDMA system. In IEEE Vehicular technology conference, May 1997.
[41] G. Xylomenos and G. C. Polyzos. Internet protocol performance over networks
with wireless links. IEEE Network, 13(4):5563, 1999.
Appendix A
Proof of Lemma 1
Lemma 1 The equation x = e
x
, with > 0, has a solution x > 0 if and only if
e. Furthermore, the solutions satisfy 1/( 1) < x < 2.
Proof : For the rst part, dene f(x) = x + e
x
. We have f(0) = 1 and f(x)
+ as x +. By continuity, f has a zero if and only if minf(x) 0. The
minimal value can be computed explicitly: f
(x) = + e
x
. If 1, f is strictly
increasing and hence positive for all x 0. So assume > 1. Then f
(x) = 0 has
a single solution at x = log . Which implies min f(x) = (log + 1) 0 if and
only if e.
For the second part, the inequality
x = e
x
> 1 + x +
1
2
x
2
implies that 0 > x
2
2( 1)x + 2 = (x ( 1))
2
+ 2 ( 1)
2
. Since e,
we have ( 1)
2
> 2. Hence any solution x must lie in the interval
|x ( 1)| <
_
( 1)
2
2
For the lower bound, use 1
1 t t/2, to get
x > 1
_
( 1)
2
2
= ( 1)
_
1
1
2
( 1)
2
_
1
1
and for the upper bound, x < 1 +
_
( 1)
2
2 < 2.
75

Full Text 01

Uploaded by

Copyright:

Available Formats

Full Text 01

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Full Text 01

Uploaded by

Copyright:

Available Formats

Automatic control in TCP over wireless

/b as the rtt corresponding to the equilibrium. The linearized

, so the poles lie outside

around the origin. Let us derive a bound > 0 for

e, and the second part gives x > 1/(

). So in this case, we have the bound > 1/(

s = z, so that s is a pole of the system. First, note that

Require to be such that e

< /2, which is always possible for some > 0.

This implies that > (1/

). So if we choose small enough, this

= {smallest positive solution , to

). The lower curve is computed

/. Figure 3.3 shows as a function

/. The convergence speed peeks for

= e, which is when the system has a

< e. To summarize, we have the following result.

. Then, the non-linear system without propagation delay

+ q/b)ds. The virtual time corresponds to the number of rtts. Then

/2. Note that

= w b. Hence, if T is suciently large, and w(t) = w

b > 0). After the outage ends, the proxy

Figure 4.3: Frame loss probability as a function of sir

occurs can be computed by conditioning on the initial state i:

will depend on the autocorrelation of e; it seems

3 1/2)(b a) > b. Hence, P

X). This means that we want to

X) corresponds to a tail of the original distribution X.

X) = c. Then, for any

X + V ), which will also depend on the robustness margin . First, dene

X+V ) will not be much larger

You might also like