Performance and security issues of TCP bulk data transfer in a last mile wireless scenario investigations and solutions

Two specific issues are addressed in this thesis:a The effect of algorithms that improve the fairness of TCP congestion avoidance on slow links and long thin networks, b The combined iss

PERFORMANCE AND SECURITY ISSUES OF TCP BULK DATA

TRANSFER IN A LAST MILE WIRELESS SCENARIO:

INVESTIGATIONS AND SOLUTIONS

VENKATESH S OBANAIK

NATIONAL UNIVERSITY OF SINGAPORE

2003

PERFORMANCE AND SECURITY ISSUES OF TCP BULK DATA

TRANSFER IN A LAST MILE WIRELESS SCENARIO:

INVESTIGATIONS AND SOLUTIONS

VENKATESH S OBANAIK

(B.Tech Electronics and Communication Engineering)

A THESIS SUBMITTED IN PARTIAL FULFILLMENT OF THEREQUIREMENTS FOR THE DEGREE OF MASTER OF SCIENCE

(COMPUTER SCIENCE)

DEPARTMENT OF COMPUTER SCIENCENATIONAL UNIVERSITY OF SINGAPORE

2003

It is often said that “It takes an artist to turn a piece of stone into a work of art” Fortunately, I had the privilege of rubbing shoulders with many such artists who

contributed in chiselling my ideas into this thesis work This thesis would not have

been possible without the:

Invaluable suggestions and encouragement of my supervisor Dr Lillykutty Jacob,Constructive criticisms and support of my co-supervisor Dr Ananda A L,Useful interactions with the research community on various mailing lists like IPv6,

Iperf, Tcpdump and Linux Users Group,Ample infrastructure and a good research environment in CIR,

Fruitful discussions with my former colleagues Saravanan, Srijith and Michael,Lively ambience and encouraging atmosphere at CIR due to my friends

Sudharshan, Rahul, Sridhar and Aurbind,Timely help by Sridhar with LATEX,

ii

Acknowledgements iii

Soothing songs on Gold 90 FM which accompanied me on lonely nights in CIR,And wonderful episodes of Seinfeld that kept me going till the end

Finally, I extend my gratitude to everybody, who in one way or the other rendered

their support and help

TCP was designed nearly three decades ago with some inherent assumptions Overthe years many fixes and solutions have been proposed to make TCP cope withchanging network conditions This research work investigates some of the proposedsolutions, studies their applicability and/or limitations in the last mile wireless sce-nario and proposes novel solutions Two specific issues are addressed in this thesis:(a) The effect of algorithms that improve the fairness of TCP congestion avoidance

on slow links and long thin networks, (b) The combined issue of performance andsecurity in a wired-cum-wireless scenario

The first part of the thesis demonstrates that fairness algorithms have a mental effect on connections traversing slow links and long thin networks Simula-tions and test-bed experiments substantiate this claim Some solutions are suggested

detri-to overcome the performance degradation

iv

1.1 Motivation 2

1.2 Research Objectives 3

1.3 Thesis Contribution 3

1.4 Thesis Organization 4

2 Background Work 6 2.1 TCP Congestion Avoidance and Control 6

2.2 Issues with TCP Congestion Avoidance and Control 7

2.2.1 Unfairness of TCP Congestion Avoidance 7

2.2.2 Inability to Identify the Nature of Loss 8

2.3 Solutions Proposed to Address the Issues 9

2.3.1 Algorithms That Improve Fairness of TCP Congestion Avoid-ance 9

2.3.2 Performance Enhancing Schemes for TCP over Wireless 11

2.4 Limitations of the Proposed Solutions 13

2.5 Summary 13

vi

CONTENTS vii

3 Fairness Algorithms and Performance Implications 15

3.1 Simulation Setup 17

3.2 Simulation Study 19

3.2.1 Behaviour of IBK, CR and CANIT Policies on Connections that Traverse Slow Links and Long Thin Networks 20

3.2.2 Impact of Last-hop Router Buffer Size on Performance 23

3.2.3 Impact of Selectively Disabling the Policies on Performance 25

3.2.4 Impact of Advertising a Limited Receive Window on Perfor-mance 27

3.3 Test-bed Experiments 29

3.3.1 Test Configuration 29

3.3.2 Impact of IBK, CANIT and CR Policies on Slow Link and LTN Connections 30

3.3.3 Impact of Last Hop Router Buffer Size on Performance 33

3.3.4 Impact of Receiver’s Advertised Window on Performance 34

3.3.5 Impact of Selectively Disabling Fairness Policies on Slow Links and Long Thin Networks 35

3.4 Recommendations 36

3.5 Summary 36

4 SPEP: Secure Performance Enhancing Proxy 38 4.1 Related Work and Issues 40

4.2 The SPEP Approach 46

4.2.1 SPEP Overview 47

4.2.2 SPEP Design Considerations 50

4.2.3 SPEP Implementation Description 52

4.3 Behavior of SPEP under Different Conditions 54

4.3.1 Presence of Packet Reordering 55

4.3.2 SPEP Mobile Handoff Scenario 56

4.4 Summary 58

CONTENTS viii

5 SPEP: Test Methodology and Performance Evalulation 59

5.1 Test Configuration 59

5.2 Performance Evaluation 60

5.3 SPEP Approach: Merits 64

5.4 Problems Encountered 67

5.5 Summary 68

6 Conclusion 69 6.1 Summary 69

6.2 Review of thesis objectives 71

6.3 Future Work 72

Bibliography 73 A Appendix I 80 A.1 Papers published related to thesis 80

List of Figures

3.1 Simulation topology 18

3.2 Congestion window variation with arrival of ACKs for slow link con-nection 21

3.3 Congestion window variation with arrival of ACKs for LTN connection 22 3.4 Goodput and loss for slow link connection 24

3.5 Goodput and loss for LTN connection 24

3.6 RTT variation for different buffer size 25

3.7 Test configuration 30

3.8 Nokia D211 PCMCIA multimode radio card 30

3.9 Variation of congestion window for slow link connection 31

3.10 Variation of congestion window for LTN connection 32

4.1 Split-Connection approach 40

4.2 Snoop approach 41

4.3 Freeze-TCP approach 43

4.4 The SPEP approach 47

4.5 IPv6 header 51

5.1 SPEP test configuration 60

5.2 Congestion window variation for LAN scenario (1 error in every 32KB) 61 5.3 Time-Sequence graph LAN scenario (1 error in every 32KB) 62 5.4 Throughput of New Reno with and without SPEP for LAN scenario 62

ix

LIST OF FIGURES x

5.5 Congestion window variation for WAN scenario (1 error in every32KB) 635.6 Time-Sequence graph WAN scenario (1 error in every 32KB) 645.7 Throughput of New Reno with and without SPEP for WAN scenario 645.8 Throughput of SPEP with NewReno V/s SPEP 65

List of Tables

3.1 Performance comparison for test configuration 1 23

3.2 Performance comparison for test configuration 2 23

3.3 Performance of slow link connection when policies are selectively dis-abled 27

3.4 Performance of LTN connection when policies are selectively disabled 27 3.5 Performance of slow link connection with a limited receive window 28

3.6 Performance of LTN connection with a limited receive window 28

3.7 Performance of slow link and LTN for various congestion avoidance policies 33

3.8 Effect of buffer size on goodput of the slow link connection 33

3.9 Effect of buffer size on goodput of the LTN connection 34

3.10 Effect of receiver window on slow link connection 34

3.11 Effect of receiver window on LTN connection 34

3.12 Performance of slow link connection when policies are selectively dis-abled 35

3.13 Performance of LTN link connection when policies are selectively dis-abled 35

4.1 SPEP loss detection and distinction algorithm 53

4.2 SPEP behavior in presence of packet reordering 55

4.3 SPEP before handoff operation 57

4.4 SPEP after handoff operation 57

xi

“The time to begin writing an article is when you have finished it to your

satisfaction By that time you begin to clearly and logically perceive what it

is you really want to say.”

- Mark Twain

1

Introduction

T he thesis studies the existing solutions for improving the performance

of TCP, investigates the applicability and/or limitations of the existingschemes in “the last mile” wireless scenario and proposes efficient solutions for it.The access network is colloquially known as “the last mile” The connection fromlocal service provider to the consumers is referred to as the “local loop” or “the lastmile” [1] Originally developed to support telephony traffic, the “local loop” of thetelecommunications network now supports both voice and Internet traffic Differentmedia can be used to provide “the last mile” connectivity, such as telephone wire,coaxial cable, fibre optics, satellite communications and wireless RF [1] Wirelessaccess to the Internet is referred to as the “last mile wireless” scenario The “last milewireless” is seen as a viable and cost-effective solution for last mile connectivity [2].Hence, it becomes essential to take a second look at the existing solutions in the

1

Chapter 1 Introduction 2

context of the last mile wireless scenario The following sections of this chapterdescribe the motivation for the research work, research objectives, contribution andthe organization of the thesis, respectively

TCP was designed nearly three decades ago and fine tuned over the years for ditional networks comprising of wired networks and fixed hosts However, the net-works have changed over the years from wired to wireless, low bandwidth to veryhigh bandwidth, stationary host to mobile host, and infrastructure based networks

tra-to ad-hoc networks Meanwhile, Internet applications have become more demandingand versatile Among today’s applications are interactive applications demanding aquick response time, bulk data transfer applications requiring high throughput andmultimedia applications sensitive to jitter Nevertheless, TCP continues to be themost widely used transport layer protocol In an attempt to equip TCP to makebetter use of the network and meet the demands of the user applications, many solu-tions have been proposed The thesis investigates two specific solutions proposed tofine tune TCP and discusses the limitations of such schemes in the last mile wirelessscenario and proposes solutions to circumvent the encountered problems

Chapter 1 Introduction 3

The thesis aims to study the existing solutions for improving the performance ofTCP and to identify specific issues for further study Investigations into the issuesconcerning the applicability and/or limitations of existing schemes in the context ofthe last mile wireless scenario and providing efficient solutions constitute the over-allobjectives of the thesis

The thesis identifies two issues with the existing schemes designed to improve theperformance of TCP: (i) The detrimental effects of fairness algorithms on the per-formance of slow links [3] and Long Thin Network [4](LTN); (ii) The limitations ofthe existing performance enhancement schemes for TCP over wireless to function in

an end-to-end IPSEC environment

Simulation and test bed experiments were conducted as part of the detailedinvestigations to study the effect of algorithms that improve the fairness of TCPcongestion avoidance on the performance of slow links and LTN Our results showthat the fairness algorithms have adverse effects on connections traversing eitherslow links or LTN We argue that it is not appropriate to apply the fairness algo-rithms for connections that traverse slow links or LTN We have studied some of thepossible solutions (increasing the last-hop router buffer size, reducing the advertised

Chapter 1 Introduction 4

window of the receiver and selectively disabling the fairness policies) in order to cumvent the adverse effects of fairness algorithms in the last mile scenario We showthat the impact can be reduced by selectively turning off the policies for slow link

cir-or LTN connections In the second part of the thesis, we present a detailed survey

of the co-existence of security and performance enhancing schemes in the last milewireless scenario We expose the limitations of the existing solutions in providingboth end-to-end security and improved transport layer performance We propose aninnovative mechanism, which we call Secure Performance Enhancing Proxy (SPEP)

to address the seemingly arduous problem of enhancing TCP performance over less networks, preserving end-to-end TCP semantics as well as ensuring end-to-endsecurity We have implemented the proposed scheme in FreeBSD 4.5 and conductedexperiments in a controlled test bed setup Our results show improved TCP per-formance in a secured environment with introduction of about 7 % overhead whencompared to the end-to-end ELN scheme in a WAN scenario with high error rates

wire-of 1 error in every 16KB wire-of data

The thesis is organized as follows Chapter 2 describes related work for enhancingthe performance of TCP data transfer in a last mile wireless scenario and iden-tifies two specific issues: (i) The effect of algorithms that improve the fairness ofTCP congestion avoidance on the performance of slow links and LTN; (ii) The com-bined issue of performance and security in a last mile wireless scenario Chapter 3

Chapter 1 Introduction 5

presents the setup for simulation and the test-bed experiments conducted to studythe effect of fairness algorithms on performance of slow links and LTN and providesrecommendations Chapter 4 presents our novel approach called SPEP which pro-vides a solution for the co-existence of IPSEC and performance enhancing solutions.Chapter 5 describes our test methodology to evaluate the performance of SPEPand presents performance results We conclude with an indication of future work inChapter 6

“To look backward for a while is to refresh the eye, to restore it, and to render

it more fit for its prime function of looking forward”

- Margaret Fairless Barber

2

Background Work

T his chapter introduces the reader to the prior work related to improving

the performance of TCP and describes the specific issues addressed bythe thesis The following sections discuss congestion control mechanisms of TCP,the issues with TCP congestion avoidance and control, and solutions proposed toaddress the issues

TCP is an end-to-end connection-oriented transport layer protocol which ensuresreliable transfer of data TCP uses a window based congestion control algorithm toreduce congestion in the network It is a self-clocking protocol and automatically

6

Chapter 2 Background Work 7

adjusts to the bandwidth of the network At the start of the connection, TCPprobes the network capacity by sending out packets at an increasingly exponential

rate This is the slow start phase and it continues until the slow start threshold is reached or a packet is lost TCP then enters the congestion avoidance phase and

sends out packets at a linear rate

Control

TCP congestion avoidance and control [5] was originally proposed by Van Jacobson

in one of the seminal papers It was proposed to solve a series of ‘congestion collapses’that occurred during 1986 The congestion avoidance and control mechanism, latersupplemented with fast recovery and fast retransmit mechanisms became the defacto standard [6] for TCP However, there are some issues with TCP congestioncontrol Two specific issues are discussed in the following subsections

2.2.1 Unfairness of TCP Congestion Avoidance

Fairness is an important criterion in the design of congestion control mechanisms.One way to define fairness is that if multiple TCP connections share a bottlenecklink, the available bandwidth is shared equally among all the connections However,

it is seen that when a bottleneck link is shared by multiple connections with shortand long round trip times (RTTs), the short RTT connections get a greater share

of the bottleneck bandwidth [7, 8] TCP uses the slow start mechanism to probe

Chapter 2 Background Work 8

the network at the start of a connection, time spent in the slow start phase isdirectly proportional to the RTT And for a long RTT connection, it means thatTCP stays in the slow start phase for a longer time when compared to a shortRTT connection This drastically reduces the throughput of short duration TCPconnections Furthermore, following each packet loss, TCP enters the congestionavoidance phase or even the slow start (in case of retransmission timeout) Duringthe congestion avoidance phase, the TCP sender increases its congestion window byatmost 1 segment after each RTT [5], thus the connections with long RTT open uptheir congestion windows relatively slower when compared to the connections withshort RTT In an attempt to counter the bias of the congestion avoidance mechanismagainst long RTT connections, and in effect, to improve the fairness, many policieshave been proposed [7, 9, 10] which will be discussed in Section 2.3.1 The policieswere designed to enable long RTT connections to open up their congestion windowsrelatively fast

2.2.2 Inability to Identify the Nature of Loss

TCP was designed for wired networks with an inherent assumption that packet losscaused by damage in the network is very small and that the loss of a packet alwayssignals congestion [5] TCP congestion avoidance and control procedures are invoked

on detection of a packet loss The occurrence of packet loss is indicated by either

a retransmission timer timeout or the receipt of duplicate acknowledgements [6,11].However, packet loss can occur for reasons other than congestion Communicationover wireless links is affected by high bit error rate, temporary disconnections, high

Chapter 2 Background Work 9

latencies and low bandwidth Losses due to bit-error rate and mobility of deviceshas a significant effect on the dynamics of TCP resulting in sub-optimal performanceand reduced throughput for the connection

In this section, we discuss the various solutions proposed to address and resolve theissues mentioned in Section 2.2.1 and Section 2.2.2

2.3.1 Algorithms That Improve Fairness of TCP Congestion

Avoidance

In an attempt to counter the bias of TCP congestion avoidance against long RTTconnections, various alternate congestion avoidance policies have been proposed.The “Constant Rate” [9] algorithm was one of the proposed solutions In this

scheme it is suggested that congestion window be increased by ‘c ∗ r2 ’ segmentsfor each RTT, where ‘c’ is some fixed constant and ‘r’ is the average round triptime In the standard congestion avoidance algorithm, the congestion window isincreased at the rate of approximately 1 segment every RTT If ‘r’ is the averageRTT of the connection, the increase in throughput of the connection would be

‘1/r’ segment/s every ‘r’ seconds This means the rate of increase in throughput is

‘1/r2’ segments/s/s Therefore the long RTT connections suffer The suggestions

in [8,9] was to modify the additive increase policy so that all connections, irrespective

of their RTT, increase their sending rate similarly Hence, it was referred to as

Chapter 2 Background Work 10

“Constant-Rate” window increase algorithm However, the choice of a proper valuefor the constant ‘c’ is an open problem According to the studies conducted in [7],the fairness properties were best at values of ‘c’ less than 100 and large values of

‘c’ made the connections very aggressive However, smaller values of ‘c’ resulted inunder utilization of the link As mentioned in [7], in reality, the choice of a propervalue for ‘c’ is not possible

“Increase-by-K” (IBK) [7] was another policy that was suggested The IBK icy was designed so that long RTT connections could increase their own throughputwithout co-operation from other connections The IBK policy suggested that thecongestion window should be increased by ‘K’ segments every RTT The policy was

pol-to be selectively enabled, only on the long RTT connections The values of ‘K’ up

to 4 was recommended for good performance [7]

Another algorithm that was proposed was Congestion Avoidance with ized Interval of Time (CANIT) [10] In the long RTT connections the arrival ofACK packets is relatively slow, compared to short RTT connections CANIT ad-dresses the fairness problem in this perspective CANIT introduces a new parametercalled Normalized Interval of Time (NIT) The congestion avoidance mechanism ismodified to increase the congestion window at the rate of RT T

Cwnd on receipt ofevery ACK Thus, all the connections increase their congestion window by the sameamount after each interval NIT CANIT with NIT value of 30ms was considered to

be most fair

Chapter 2 Background Work 11

2.3.2 Performance Enhancing Schemes for TCP over

to address the problem The link-layer protocols employ two classes of techniques (i)error correction using forward error correction (FEC), and (ii) retransmission of lostpackets using automatic repeat request (ARQ) The link-layer protocols for digitalcellular techniques CDMA and TDMA primarily use ARQ The AIRMAIL [12]protocol uses a combination of FEC and ARQ for loss recovery

B Split Connection Schemes

Split connection schemes attribute the performance degradation of TCP over less to the inability of TCP to cope with the dynamics of wireless link The splitconnection schemes as the name suggests, splits the TCP connection from sender

wire-Chapter 2 Background Work 12

to receiver at the base station referred to as the mobility support node One nection is established between the sender and the base station and the other fromthe base station to the receiver Split connection schemes such as I-TCP [13]useregular TCP for its connection over the wireless link, while other schemes such asMTCP [14], recommend a protocol optimized for wireless links to be used for theconnection between the base station and the receiver

con-C End-to-End Schemes

End-to-end schemes abide by the principle that end-to-end argument is one of thearchitectural principles in the design of the Internet Hence, all problems withTCP have to be solved end-to-end The schemes which maintain the end-to-endsemantics are Freeze-TCP [15], TCP HACK [16], SACK [17] Freeze-TCP proposes

an end-to-end solution to enable TCP to cope with long periods of disconnectiondue to degraded wireless link The TCP receiver in Freeze-TCP advertises a zerowindow in case of an imminent link failure The sender reacts to the zero windowadvertisement by freezing all retransmit timers and entering persist mode TCPHACK [16] proposes a scheme, in which the receiver can distinguish between thenature of loss congestion or corruption The information about the nature of loss isconveyed to the sender The sender retransmits the packets lost due to corruptionwithout invoking congestion control while the packets lost due to congestion arehandled normally SACK [17] provides a mechanism which enables the TCP sender

to recover from multiple losses in a window of data transmitted The SACK enabledreceiver uses the TCP SACK option to acknowledge the blocks of data received in

Chapter 2 Background Work 13

sequence The sender retransmits the lost segments, thereby reducing the number

of probable retransmission timeouts

D Other Schemes

The Snoop protocol [18] proposed by Hari et al is another performance enhancementscheme, which was designed with the intent that local problems should be solvedlocally Therefore, Snoop suppresses the duplicate acknowledgements that signal loss

in wireless link and locally retransmit the lost segments Thereby, Snoop achieves aremarkable improvement in the performance of TCP over wireless links

The solutions proposed in Section 2.3 overcome the shortcomings of TCP mentioned

in Section 2.2 However, the solutions take a myopic view of the problem Althoughthe proposed solutions attempt to resolve the issue at hand, they have some limita-tions which restrict their applicability In the following chapters, the drawbacks areexposed and solutions are proposed to counter the same

This chapter presents a brief description of TCP congestion avoidance and controland the issues concerning it Two specific issues with TCP congestion avoidance andcontrol mechanism are identified for further examination, namely: (i) the unfairness

Chapter 2 Background Work 14

of TCP Congestion Avoidance, and (ii) the inability to identify the cause of packetloss Various solutions suggested to overcome these issues are described Finally, it

is mentioned that though the proposed solutions address and resolve the problem

at hand, they may not be applicable for all scenarios

“I didn’t think; I experimented.” - Wilhelm Roentgen

3

Fairness Algorithms and Performance

Implications

T his chapter reports the simulation studies and test bed experiments

con-ducted in connection with the first issue addressed by this thesis Theobjective of the experiments is to show that although the fairness algorithms men-tioned in Section 2.3.1 address and resolve the issues highlighted in Section 2.2.1,they tend to be harmful to connections traversing slow links and long thing networks

Currently, there are a significant number of users connecting to the Internetthrough slow last-hop links, usually the 56Kbps modem links And now, with theincrease in the number of wireless devices, there is a growing number of users con-necting to the Internet through Wireless WANs (W-WAN), often known as the LongThin Networks (LTN)

15

Chapter 3 Fairness Algorithms and Performance Implications 16

In today’s Internet, a 56Kbps modem link typically used by a laptop to accessthe Internet is considered as a slow link One way propagation delay on slow links isaround 50ms The Long Thin Networks (LTNs) are called “Long” networks as theyare characterized by large round trip times and are “Thin” since the networks usuallyhave a low bandwidth The common examples of LTNs are the Wireless WANs (W-WAN) like the GSM (Global System for Mobile communications), CDPD (CellularDigital Packet Data) and Ricochet [4] For a W-WAN like Ricochet the typicalRTT value is around 500ms and the bandwidth is 24Kbps [4] As can be seen, bothLTNs and slow links are low speed links However, unlike slow links, LTNs haverelatively long and highly varying RTTs, which results in variable bandwidth delayproducts Some LTNs like GSM offer a reliable wireless link by employing link levelerror recovery mechanisms Therefore, increased link error rate would also result inincreased RTT

The slow links and LTNs are usually considered to be the last-hop links, necting end user to the global Internet That is, slow links or the LTNs are used

con-as access links, with most part of the connection traversing through relatively highspeed links Hence the last-hop router to which the slow link or LTN is connected,may receive packets at a higher speed than it can forward on a low speed link Thelast-hop router usually has a limited number of buffers configured for each outboundlink Sometimes even as low as a buffer size of 3 packets [3,4,19] The last-hop routerbecomes a point of congestion and excess packets get dropped This is an existing is-sue and many studies have been conducted [19–21] to see the implications of limitedbuffer size on the last-hop router

Chapter 3 Fairness Algorithms and Performance Implications 17

As described above, it is not uncommon to have slow “access” links connecting

to the Internet We consider the following cases : (a) a slow link like a 56Kbpsmodem link used as an access link to connect to the global Internet and is a part

of a long RTT connection; (b) a connection which has a long RTT because of thepresence of a LTN as an access link in the path In both the cases the TCP sendercan only perceive a long RTT connection, but it is ignorant of the characteristics

of the paths in the network When the TCP sender is equipped with the policiesproposed in [7,9,10] that are designed to improve fairness, it reacts by opening up thecongestion window at a much higher rate However, the connections described in thecases above only offer a long RTT but not high bandwidth Hence, in these cases,the increased probing harms the connection As mentioned earlier, the policies

of [7, 9, 10] were designed so that long RTT connections get a fair share of thebottleneck bandwidth However, these policies are not appropriate for slow linksand LTNs, since they themselves are the bottlenecks in the network In the followingsections we evaluate the impact and identify ways to reduce the impact of the policiesproposed in [7, 9, 10] on connections like the ones described above

We use Network Simulator “ns” [22] to evaluate the impact of the proposed policies

on slow links and LTNs In this section, we describe the topologies used in oursimulation study The test configuration shown in Figure 3.1 was chosen as it hasbeen used previously by the research community [7, 10, 23] in studies related to the

Chapter 3 Fairness Algorithms and Performance Implications 18

improvement of fairness All the links unless specified have a bandwidth of 10Mbps.The path from source 3 to sink 3 represents a long RTT connection comprising aslow last-hop link of bandwidth 56Kbps and one way propagation delay of 50ms.Similar values were used in earlier studies to emulate a dial up access link [24] The

Last−hop router

3

1.5Mbps 200ms

150ms

50ms

5ms

5ms 5ms

5ms

400ms

100ms

1.5Mbps 25ms

4 Sink

Sink 5

Sink 2

Sink 1 Src1

(a) Test configuration 1 : Slow link

Last−hop router

3

1.5Mbps 200ms

150ms

50ms

5ms

5ms 5ms

5ms

400ms

100ms

1.5Mbps 25ms

4 Sink

Sink 5

Sink 2

Sink 1 Src1

(b) Test configuration 2 : LTN

Figure 3.1: Simulation topology

bandwidth and propagation delay of the path connecting router R2 and the last-hoprouter are chosen to be 1.5Mbps and 25ms, respectively The values were chosen inaccordance with a study previously conducted on the last-hop router with limitedbuffer size [19] We consider RTT of each connection to be twice the sum of all the

Chapter 3 Fairness Algorithms and Performance Implications 19

link delays along the path The utilization of the link connecting router R2 andthe last-hop router is kept close to 1 In the test configuration 2, the connectionfrom source 3 to sink 3 represents a connection that traverses a long thin networkwith bandwidth 9.6Kbps and a constant one way propagation delay of 200ms Wecall the connection that traverses slow link (resp LTN) as the Slow link connection(resp LTN connection)

First, our interest is to see the possible impact caused by the CR, IBK and CANITpolicies [7, 9, 10] on connections traversing slow links and LTN Hence, we try toanalyze the behaviour of those policies in the following subsection A buffer size

of 3 packets, on the last-hop router is a common configuration followed by InternetService Providers (ISPs) We have chosen a buffer size of 3 packets in our stud-ies since it is a common case [3, 4, 19] The MTU (Maximum Transmission Unit)size is chosen as 296 bytes, in accordance with the recommendations in [3] Weconducted 20 independent simulation runs for each experiment and the duration ofeach simulation was 100s

We evaluate the performance of the connections via two metrics, namely, put and packet losses incurred These metrics enable us to identify the data thatactually gets across to the receiver and the amount of data lost due to buffer over-flow In “subsection B”, we study the impact of those policies by varying the buffer

good-Chapter 3 Fairness Algorithms and Performance Implications 20

sizes from 3 to 40 packets and try to identify the ways to reduce the impact on theslow links and LTNs

3.2.1 Behaviour of IBK, CR and CANIT Policies on

Con-nections that Traverse Slow Links and Long Thin works

Net-We are interested in seeing the amount of increase in congestion window, caused byeach arriving ACK when different policies are employed.Figures 3.2 (a) (b) and (c)are plots of congestion window against the ACKs received, for slow link connection(test configuration 1) when CANIT, CR and IBK policies are used, respectively.IBK policy has to be selectively enabled on long RTT connections Therefore, inall our experiments we enable the IBK policy only on the connections from source

3 to sink 3, source 4 to sink 4 and source 5 to sink 5 All the policies are comparedwith the behavior of the standard congestion avoidance algorithm Figures 3.3 (a)and (b) are plots of congestion window against ACKs received, obtained by usingtest configuration 2 Fig 3.3 (a) shows the CR policy compared with the standardpolicy Fig 3.3 (b) shows the variation in congestion window caused by the IBKpolicy compared with standard policy During slow start all the schemes behavesimilarly, which is expected However, during the congestion avoidance phase, thestandard algorithm reacts to congestion by increasing its congestion window by

1/cwnd on the receipt of every acknowledgement, while the other policies increase

their congestion window at a much higher rate In the congestion avoidance schemesthat consider RTT as a parameter to increase the congestion window, there is a steep

Chapter 3 Fairness Algorithms and Performance Implications 21

N IT · 1

cwnd on the receipt of every acknowledgement

We have considered the optimum value of NIT to be 30ms, as mentioned in [10].The RTT of the connection could be as high as 500ms Therefore, this algorithmwould enable the TCP sender to inject many segments into the network on thereceipt of every ACK On connections with slow links or LTNs and limited last-hoprouter buffers, this could be very harmful Figures 3.2 and 3.3 show a large number

of retransmission timeouts, and congestion window being reset to 1 The effect is

Chapter 3 Fairness Algorithms and Performance Implications 22

(b) Constant Rate Vs Standard

Figure 3.3: Congestion window variation with arrival of ACKs for LTN connectionreflected in the number of losses and goodput of the connection as can be seen inTable 3.1 and Table 3.2

In Constant Rate Policy, the congestion window is increased at the rate of

c·RT T ·RT T

cwnd on the receipt of every acknowledgement According to [7], values of

’c’ less than 100 were best but smaller values of ’c’ decreased the link utilization.Therefore, in our studies we consider the values of ’c’ as 10 Higher values of c wouldonly make the connection more aggressive [7] The losses are relatively high and thegoodput is lower than that achieved by the standard algorithm The CR policy,even with a value of ’c’ as small as 10, could prove to be very aggressive when theRTT is high, like in the cellular links The impact is worse if a higher value of ’c’ ischosen, since it would cause more packets to be injected into the network In testconfiguration 2, we have assumed the LTN to have a typical value of 200ms for thepropagation delay However, in reality LTNs are known to have, long and varyingRTTs, sometimes in the order of seconds As can be seen in the above policies,higher values of RTT will only make the connection more aggressive, since more

Chapter 3 Fairness Algorithms and Performance Implications 23

packets will be injected into the network

Policy Goodput (Kbps) Losses

Table 3.1: Performance comparison for test configuration 1

Policy Goodput (Kbps) Losses

Table 3.2: Performance comparison for test configuration 2

It is seen that the losses incurred is relatively less in the case of ”Increase-by-K”policy According to the policy, congestion window is increased by a minimum of(K/cwnd, 1 segment) on the receipt of every acknowledgement The limit ensuresthat the connections can only get as aggressive as slow start As can be seen inthe graph in Figures 3.2 (c) and 3.3 (a), the slope of IBK during the congestionavoidance phase resembles that of slow start

3.2.2 Impact of Last-hop Router Buffer Size on Performance

We have seen in the previous subsection that the RTT-aware fairness policies sendout segments in bursts and buffer size of 3 packets for the last hop router is toosmall to absorb the bursts In this subsection, we study the effect of having in-creased last-hop router buffer sizes Figures 3.4 (a) and (b) are obtained using thetest configuration 1, and depict the variation in goodput and losses respectively, on

Chapter 3 Fairness Algorithms and Performance Implications 24

varying the buffer size Figures 3.5 (a) and (b) are obtained by using Test uration 2 and depict the variation in goodput and losses incurred respectively, onvarying the buffer size Degradation in goodput and increased losses are seen,

(a) Goodput

0 50 100 150 200 250 300 350

(a) Goodput

0 10 20 30 40 50 60 70

(b) Loss

Figure 3.5: Goodput and loss for LTN connection

especially when the buffer sizes at the last-hop router are small Large variations

in goodput is seen, as the buffer sizes are increased Sometimes, it is seen that thegoodput drops as the buffer size is increased This happens when losses occur at theend of a connection, causing a retransmission timeout Similar observations wererecorded in a study conducted previously [21] The losses incurred due to the use

Chapter 3 Fairness Algorithms and Performance Implications 25

of the proposed policies are more than 2 to 3 times than that incurred by using thestandard policy As expected, the standard congestion avoidance policy performsmuch better with losses being very low and the goodput relatively high With largebuffer sizes, the traffic bursts can be absorbed and the use of the proposed policies

do not have much impact But increased buffer sizes introduces additional queuingdelay and therefore increases RTT of the connection The effect of increased buffersize on RTT for a LTN connection is shown in Figure 3.6 The RTT values for buffersizes of 3, 7 and 15 are plotted We see that the value of RTT is the least at a buffersize of 3

5 10 15 20 25 30

Figure 3.6: RTT variation for different buffer size

3.2.3 Impact of Selectively Disabling the Policies on

Perfor-mance

In the test configuration 1, the bottleneck link for the connection between source 3and sink 3 is 56Kbps, whereas for all the other connections the bottleneck link is1.5Mbps The proposed policies [7,9,10] are designed to ensure a fair allocation of the

Chapter 3 Fairness Algorithms and Performance Implications 26

shared link (in this case, 1.5Mbps) to all the competing connections However, it can

be seen that irrespective of the policies used to improve fairness, the connection fromsource 3 to sink 3 can only have a maximum throughput of 56Kbps It is the samecase for the connection with LTN, which can only a have a maximum throughput of9.6Kbps Hence it is not meaningful in trying to get a fair allocation of the sharedlink for connections with slow links and LTNs The policies are not appropriate tosuch connections and need not be applied We have seen in earlier subsections, thatsuch policies which are intended to provide a fair share will only end up harming theconnections Therefore, in the next experiments we selectively disable the policies forconnections traversing slow links or LTNs In other words, the standard congestionavoidance policy is used on connections traversing slow links or LTNs, while all theother connections use the alternate policies, namely IBK, CR or CANIT In Tables3.3 and 3.4, column 3 indicates the goodputs for the Slow Link/LTN connectionwhen the corresponding policy is disabled for the Slow link / LTN connection ( i.e,when the the standard policy is used on the slow link/LTN connection while all otherconnections are enabled with CR, IBK or CANIT policies) As mentioned earlier,IBK policy is selectively enabled for long RTT connections We use a limited buffersize of 3 packets at the last-hop router We notice that the standard policy gives thebest performance for slow link and LTN connections However, as we have seen inTable 3.1, with standard policy being enabled on all the connections, the throughput

of the slow link connection is 16.54Kbps which is greater than the values in column

3 of Table 3.3, when competing connections use aggressive congestion avoidanceschemes The same applies to the goodput of the LTN connection The goodput of

Chapter 3 Fairness Algorithms and Performance Implications 27

the LTN connection when all the competing connections use the standard algorithm

is 7.62 Kbps as in Table 3.2, which is greater than the values in column 3 of Table3.4 This is expected since the competing connections are enabled with alternatepolicies and can open up their windows at a much higher pace compared to theconnections using standard congestion avoidance policy

Goodput (Kbps)Policy Enabled on all connections Disabled only for slow link

Table 3.4: Performance of LTN connection when policies are selectively disabled

3.2.4 Impact of Advertising a Limited Receive Window on

Performance

The recommendations in [3] suggest that the hosts which are directly connected

to low speed links could advertise a limited receive window in order to reduce oreliminate the losses at the last-hop router In the next set of experiments, instead

of selectively turning off the policies for the slow link and LTN connections, weallow the congestion window to increase according to IBK, CR or CANIT policiesbut restrict the amount of data injected into the network by advertising a smallerreceive window in accordance with the recommendations in [3] The advertised

Chapter 3 Fairness Algorithms and Performance Implications 28

receive window is chosen to have a value slightly more than the pipe capacity ofthe low speed link inclusive of the last-hop router buffer size In a similar studyconducted on slow wireless links [20], it was found that a limited advertised receivewindow of 2KB was optimum Hence, we have chosen the value of 2KB in ourstudies As we can see in Table 3.5 and Table 3.6, there is a marked decrease inthe number of losses incurred And for the connection using CANIT, the goodput

is increased since the number of losses incurred is vastly decreased However, there

is a marginal decrease in the goodput in other cases, in spite of decreased losses

A smaller advertised receive window restricts the amount of data injected into thenetwork by the TCP sender Therefore, the number of packet losses incurred due

to buffer overflows at the last-hop router is reduced However, imposing a limit onamount of data transmitted, also limits the sender from transmitting new segmentsduring the fast recovery phase [21] Besides, the competing connections are enabledwith aggressive policies These are the reasons for the decrease in goodput