Improving Performance in Heterogeneous Networks: A Transport Layer Centered Approach

(1)

$IVISION

$EPARTMENT

+ARLSTAD

*OHAN

)MPROVING (ETEROGENEOUS

(2)

+ARLSTAD

*OHAN

)MPROVING (ETEROGENEOUS

(3)

*OHAN

#ENTERED

$ISSERTATION

+ARLSTAD )33.

)3".

$ISTRIBUTION

+ARLSTAD

$IVISION

$EPARTMENT 3%

37%$%.

WWWKAUSE

0RINTED

(4)

Improving Performance in Heterogeneous

Networks: A Transport Layer Centered Approach

JOHAN GARCIA

Department of Computer Science, Karlstad University

Abstract

The evolution of computer communications and the Internet has led to the emergence of a large number of communication technologies with widely different capabilities and characteristics. While this multitude of technologies provides a wide array of possibilities it also creates a complex and heterogeneous environment for higher-layer communication protocols. Speciﬁc link technologies, as well as overall network heterogeneity, can hamper user-perceived performance or impede end-to-end throughput. In this thesis we examine two transport layer centered approaches to improve performance.

The first approach addresses the decrease in user satisfaction that occurs when web waiting times become too long. Increased transport layer flexibility with regards to reliability, together with error-resilient image coding, is used to enable a new trade- off. The user is given the possibility to reduce waiting times, at the expense of image fidelity. An experimental examination of this new functionality is provided, with a focus on image-coding aspects. The results show that reduced waiting times can be achieved, and user studies indicate the usefulness of this new trade-off.

The second approach concerns the throughput degradations that can occur as a consequence of link and transport layer interactions. An experimental evaluation of the GSM environment shows that when negative interactions do occur, they are coupled to large variability in link layer round-trip times rather than simply to poor radio conditions. Another type of interaction can occur for link layers which expose higher layers to residual bit errors. Residual bit-errors create an ambiguity problem for congestion controlled transport layer protocols which cannot correctly determine the cause for a loss. This ambiguity leads to an unnecessary throughput degradation. To mitigate this degradation, loss differentiation and notiﬁcation mechanisms are proposed and experimentally evaluated from both performance and fairness perspectives. The results show that considerable performance improvements can be realized. However, there are also fairness implications that need to be taken into account since the same mechanisms that improve performance may also lead to unfairness towards ﬂows that do not employ loss differentiation.

Keywords: transport protocols, partial reliability, image transfer, transcoding, loss dif- ferentiation, fairness, wireless networks, network emulation

i

(5)

(6)

Acknowledgments

The person who helped me start my journey into the land of academic research is the same that has stood by me during the years with help and guidance. I am forever grateful to my supervisor Anna Brunstr¨om for this.

My colleagues and co-authors in the DISCO group have also been a great help. Thanks Stefan Alfredsson, Katarina Asplund, Karl-Johan Grinnemo, Sean Schneyer, and Annika Wennstr¨om. Thanks also to all colleagues at the Computer Science department for making it such a great place to work.

During my stay at the University of Delaware I was well taken care of by professor Paul Amer, Armando Caro, and Janardhan Iyengar. Thanks for making my stay memorable.

Several of the research projects I have taken part in has had reference groups with exter- nal experts. I thank all reference group participants for their input and helpful comments.

Finally, I thank my Louise and our kids Emmy and Max for always being there, and for providing a good life outside of work.

With gratitude I acknowledge the ﬁnancial support provided by the KK-foundation, Ericsson, CMIT, Tieto-Enator, Telia and the NEWCOM EU-project.

iii

(7)

(8)

List of Appended Papers

This thesis includes the following 8 papers. References to the papers will be made as Paper I, Paper II etc.

I. J. Garcia, A. Brunstrom, and K. Asplund. Application and Transport Layer Flex- ibility - A User-inﬂuenced Web Browsing System. Under submission.

II. J. Garcia and A. Brunstrom. Impact of Partial Reliability on Web Image Transport Performance. In Proceedings PromoteIT 2002, Sk¨ovde, Sweden, April 2002.

III. J. Garcia. JPEG Transcoding - Efﬁciency and Robustness Aspects. Karlstad Uni- versity Studies 2002:10, Karlstad University, Sweden.

IV. A. Wennstr¨om, A. Brunstrom, J. Garcia, and J. Gustafsson. GSM/GPRS Testbed for TCP Performance Evaluation. In Proceedings International Workshop on Wired/Wireless Internet Communications (WWIC 2002), Las Vegas, USA, June 2002.

V. J. Garcia and A. Brunstrom. Checksum-based Loss Differentiation. In Proceed- ings 4th IEEE Conference on Mobile and Wireless Communications Networks (MWCN 2002), Stockholm, Sweden, September 2002.

VI. J. Garcia and A. Brunstrom. Evaluation of Transport Layer Loss Notiﬁcation in Wireless Environments. In Proceedings International Conference on Network- ing (ICN2005), Boucan Canot, France, April 2005. Lecture Notes in Computer Science (LNCS), Springer-Verlag.

VII. J. Garcia and A. Brunstrom. An Experimental Study on the Performance and Fair- ness of Loss Differentiation for TCP. In Proceedings International Conference on E-Business and Telecommunication Networks (ICETE 2004), Setubal, Portugal, August 2004.

VIII. J. Garcia, S. Alfredsson, and A. Brunstrom. The Impact of Loss Generation on Emulation-based Protocol Evaluation. Under submission.

Minor editorial changes have been made to some of the papers. Paper II was changed to use the term “latency reduction” instead of “throughput gain” to improve consistency with paper I.

v

(9)

Comments on my participation

For paper I, my contribution consists of writing the parts on image coding and system experiments, as well as constructing the distributed application used for performing the user experiments. I am responsible for the ideas, implementations, experiments and results, and writing for papers II, III, V, VI, and VII. My co-author has contributed with discussions and style of writing. For paper IV, part of the data collection software was written by me and I contributed with discussion and style of writing. For paper VIII, the experiments and all writing except the background section is my work. The implementation was joint work between the co-authors.

Other Work

In addition to the included publications I have also authored or co-authored a number of additional publications. Overlap exist between the included publications and the publications in this list. Most noteworthy, Paper II is an extended version of Paper 10 and Paper III is a collection and extension of Papers 3, 6, and 7.

1. K. Asplund, A. Brunstrom, J. Garcia, K-J. Grinnemo, and S. Schneyer. PRTP - a Partially Reliable Transport Protocol for Multimedia Applications: Background Information and Analysis. Karlstad University Studies 99:05, Karlstad University, Sweden, 1999.

2. S. Schneyer, K. Asplund, A. Brunstrom, and J. Garcia. PRTP: A Partially Reliable Transport Protocol for Multimedia Applications. In Proceedings International Symposium on Intelligent Multimedia and Distance Education (ISIMADE 1999), Baden-Baden, Germany, August 1999.

3. J. Garcia and A. Brunstrom. Efﬁcient Image Transfer for Wireless Networks. In Proceedings 2nd International Conference on Advanced Communication Technol- ogy (ICACT 2000), Muju, South Korea, February 2000.

4. A. Brunstrom, K. Asplund and J. Garcia. Enhancing TCP Performance by Al- lowing Controlled Loss. In Proceedings SSGRR 2000 Computer & eBusiness Conference, L´Aquila, Italy, 2000.

5. K. Asplund, A. Brunstrom and J. Garcia. Decreasing Transfer Delay Through Partial Reliability. In Proceedings Protocols for Multimedia Systems (PROMS 2000), Cracow, Poland, October 2000.

6. J. Garcia and A. Brunstrom. A Robust JPEG Coder for a Partially Reliable Transport Service. In Proceedings 7th International Workshop on Interactive Distributed Multimedia Systems and Telecommunication Services (IDMS 2000), Enschede, The Netherlands, October 2000, Lecture Notes In Computer Science (LNCS), Springer-Verlag.

vi

(10)

7. J. Garcia and A. Brunstrom. Progressive Parsing Transcoding of JPEG Images. In Proceedings 7th International Workshop on Mobile Multimedia Communications (MoMuC 2000), Tokyo, Japan, October 2000.

8. A. Wennstr¨om, J. Garcia, J. Gustavsson and A. Brunstrom. TCP and Link Layer Interactions: Implications for the Wireless Internet. In Proceedings IEEE Vehicu- lar Technology Conference (VTC Spring 2001), Rhodos, Greece, May 2001.

9. J. Garcia and A. Brunstrom. A Robust JPEG Coder for a Partially Reliable Trans- port Service. in Proceedings PromoteIT 2001, Ronneby, Sweden, April 2001.

10. J. Garcia and A. Brunstrom. An Experimental Performance Evaluation of a Par- tially Reliable Web Application. In Proceedings International Network Confer- ence 2002 (INC 2002), Plymouth, Great Britain, July 2002.

11. J. Garcia and A. Brunstrom. Explicit and Implicit Loss Notiﬁcation for Error- prone Links. In Proceedings PromoteIT 2003, Visby, Sweden, May 2003.

12. J. Garcia and A. Brunstrom. Transport Layer Loss Differentiation and Loss Noti- ﬁcation. In Proceedings First Swedish National Computer Networking Workshop (SNCNW 2003), Arladastad, Sweden, September 2003.

13. J. Garcia and A. Brunstrom. Effects of Loss Notiﬁcation and Loss Differentia- tion on DCCP and TCP Performance. In Proceedings PromoteIT 2004, Karlstad, Sweden, May 2004.

14. K-J. Grinnemo, J. Garcia, and A. Brunstrom. Taxonomy and Survey of Retrans- mission Based Partially Reliable Transport Protocols. Computer Communica- tions, Elsevier, 2004. Vol. 27. Issue 15. pp. 1441-1452.

15. J. Garcia and A. Brunstrom. Transcoding of Image Data. Chapter in Perspectives of Multimedia: Communication, Media and Information Technology. Burnett, Brunst¨om, Pettersson (eds), Wiley, 2004.

16. J. Garcia and A. Brunstrom. Issues with TCP-Friendly Rate Control in DCCP. In Proceedings Second Swedish National Computer Networking Workshop (SNCNW 2004), Karlstad, Sweden, November 2004.

vii

(11)

(12)

Introductory Summary

(15)

(16)

1. Introduction 3

1 Introduction

Today, a typical workplace computer can exchange data with any of the hundreds of mil- lions of other Internet-connected computers around the world. In a split second, client software on a local computer can communicate with a remote server to allow users to gather information, exchange mails, perform business transactions or a multitude of other activities. What makes all this possible is the evolution of computer networking during the last 40 or so years. Computers were initially solitary, central, resources providing computing power to a few expert users. The advent of time-share access then helped to create a wider need for computer communications. Research sponsored by the US Department of Defense raised the interest in packet switching in the 60s, since it allowed for networks without single points of failure. In 1974, Cerf and Kahn pub- lished their seminal paper [19] on the interconnection of networks. The ideas proposed evolved into the Internet Protocol (IP) [58], the Transport Control Protocol (TCP) [59], and other protocols. This set of protocols, referred to as the TCP/IP protocol suite, are the underpinnings of the global, large-scale Internet of today.

The inherent complexity of large and complex systems such as global computer networks creates a need for good abstractions. A layering approach allows partitioning of functionality, supports enhanced modularity, and provides a helpful abstraction when reasoning about network systems. The de facto functionality of the different protocols in the TCP/IP protocol stack, in combination with the usefulness of layering, led to the emergence of the four layer TCP/IP reference model [46]. An alternate protocols stack, the ISO/OSI stack, was developed some years after the TCP /IP stack. In conjunction with the ISO/OSI protocol stack, the seven-layer OSI reference model [72] was constructed. Although the ISO/OSI protocol stack did not have a lasting impact the OSI reference model has, together with the TCP/IP reference model, shaped thinking and reasoning in both tele- and computer-communications for some time. Both these reference models are shown in Figure 1. Although the two models deviate slightly in how they partition the required functionality into different layers, they share the basic ideas.

In order to achieve communication, a layer uses services provided by the layer below it. Between each pair of adjacent layers there is an interface. Interfaces describe the services that are available, and specify how user-data and control information can be exchanged between layers.

Our research has the overall goal of examining the potential for improved communications performance by refining the functionality, flexibility and information exchange of different layers. Heterogeneous networks, which may have challenging characteristics such as low bandwidth, large delay variations or high error rates, provide interesting target environments for this work. From a general viewpoint, this work examines how communications performance in heterogeneous networks can be improved by increased flexibility at the transport layer together with an increased flow of information between layers.

(17)

4 Introductory Summary

Not present in the model

Network Transport Application

Host−to−network

1 Physical

Data link Network Session Presentation

Transport Application

2 3 4 5 6 7

OSI TCP/IP

Figure 1: The OSI and TCP/IP reference models, from [67]

The general perspective of transport layer flexibility has in this thesis been examined in relation to two other layers, the application layer and the link layer. Each of these two layers is covered in a part of the thesis. The first part is concerned with the interaction between the application and transport layers. In this case, additional flexibility in the transport layer allows a more flexible transport service to be provided to the application layer. There exists several aspects of the transport service that can be modified in order to increase flexibility. The work in this thesis concentrates on one specific aspect, increased flexibility with regards to reliability. We extend the service interface so that an application can communicate its reliability requirements to the transport layer. Adequate mechanisms in the transport protocol adapt their behavior in response to this information, and the end result is a transport protocol which can provide a more flexible service with regards to reliability. To evaluate the effects of this specific increase in flexibility, we examine one specific application. The chosen application is image transfer in a web browsing context. This application requires the use of a robust image coder, and the work of designing and implementing such a coder is also described.

The second part of this thesis concerns interactions between the link layer and the transport layer. The first aspect examined is the potential for negative interactions between the retransmission mechanisms at the link and transport layers. This aspect is examined experimentally for a GSM link by using actual protocol implementations and real GSM network hardware together with an emulated radio environment. The second aspect examined is how a more flexible congestion control behavior at the transport layer can improve performance over links with residual bit-errors. In this case, information from lower-layer checksumming operations are forwarded to the transport layer to allow the transport layer to infer the cause of a packet loss. This information can then be acted on by the transport layer to provide a more appropriate congestion behavior for link-related losses. Design, performance and fairness issues of checksum-based loss differentiation and the related loss notification mechanisms are examined.

(18)

2. Application and Transport Layer 5

The remainder of this introductory summary is as follows. The next section gives an overview to application and transport layer interactions and discuss related work. The following section provides an overview on transport and link layer interactions and also cover related work. Then, research issues are discussed, followed by a section outlining the contributions of this work. A summary of the appended papers is then provided, followed by a ﬁnal section which provides the conclusions and an outlook on future work.

2 Application and Transport Layer

2.1 Overview of Area and Thesis Work

Being the top-most layer, the application layer interfaces with a human end-user via an application. In addition to handling the application data that are to be communi- cated using the services of the transport layer, the application layer can potentially also provide layers below it with user specific information about preferences, or other meta- information that can be used to adapt the transport or lower layers. While there exists a multitude of applications and application protocols in the Internet today, there are only two transport protocols in widespread use. These two major transport protocols on the Internet, UDP [57] and TCP [59], provide very different transport services to the application layer, and have only a limited amount of flexibility. UDP provides unordered, unreliable datagram delivery without flow control or congestion control. TCP provides ordered, reliable byte stream transfer with flow and congestion control. The services provided by these two protocols are quite restricted in relation to the heterogeneity of application requirements. One approach to address this lack of flexibility has been for the application to implement additional functionality on top of UDP, often within the framework of the Real Time Protocol (RTP) [63]. Although the application in this case can tailor the transport layer behavior, this approach adds considerable complexity to the application layer and becomes heavily tied to the specific application. Recent transport layer protocols such as the datagram congestion control protocol (DCCP) [41] and the stream control transfer protocol (SCTP) [65] have been developed to provide new transport services and additional flexibility. DCCP provides an unreliable delivery of datagrams, and has a flexible mechanism for using different congestion control schemes.

SCTP provides reliable and congestion controlled delivery of messages, but it also supports several new concepts such as multi-homing and multi-streaming. SCTP also has provisions for providing a more flexible transport service by allowing partial order, and also partial reliability [66]. The work done in this thesis examines flexible transport services with a focus on reliability. To be able to examine the effects of partial reliability in relation to a well known protocol, we designed and implemented a partially reliable transport protocol (PRTP) as a receiver-based modification to TCP. PRTP could then be used to explore the potential benefits of partial reliability.

(19)

In order to evaluate our PRTP protocol in an application context we developed a suitable Internet image transfer application. The relevance of efﬁcient image transfer is evident as the rapid growth of the Internet has to a large extent been fueled by one application: the World Wide Web (WWW). An important part of the WWW is the images embedded in web pages. Efﬁcient transfer of images over the Internet is dependent on effective image coding to reduce the amount of data needed to represent an image.

Although there are coders available for recent standards such as JPEG 2000 [27] and PNG [13], the major image coding standards currently in use on web-pages are the older JPEG [38] and GIF [31]. JPEG is best suited for photo-like images and GIF ﬁts better for logotypes, small images, “graphic” text and images with relatively few colors. The work described in this thesis tries to increase the image transfer performance by utilizing the speciﬁc reliability versus latency trade-off made possible by PRTP. Paper I gives an overview of the different parts of the web image transfer application that was developed, and also presents both system level experiments and a user study. Paper II focuses on system level experiments and examines a part of them in greater detail.

Out of the two major coding standards, we chose to concentrate the image coding effort on JPEG. We developed a robust JPEG coder which can be used together with a partially reliable transport service. In the context of image transfer, the reliability versus latency trade-off maps to a trade-off between image quality and latency. By using the PRTP protocol a minimum reliability can be guaranteed, corresponding to enforcing a minimum image quality. To handle packet losses our robust JPEG coder utilizes extended restart markers to be able to resynchronize effectively, interleaving to spread the losses, and error concealment to hide the losses. In addition we have also incorporated fast compression level transcoding into our coder. The image coding related parts of this work is presented in Paper III.

We believe that increased application and transport layer flexibility makes it easier for applications to be useful in a large number of heterogeneous environments, since some trade-offs need not be fixed at the time of design of the application and/or transport layer protocol, but rather can be adjusted according to current conditions and user preferences. However, this increase in flexibility also increases the implementation complexity, an issue which must also be considered. In the case of PRTP, it can be realized with minor receiver-based modifications to TCP, and a simple extension to the socket interface is used to communicate the extra cross-layer information regarding reliability requirements.

2.2 Related Work

Considerable earlier work has been done with regards to protocols which provide a partially reliable service. Dempsey studied the concept of partial reliability in his thesis [25]. He examined two different error control schemes, slack ARQ and PECC, that are based on time limited retransmissions in the context of interactive voice. In [2] Par- tially Ordered Connection (POC) is described. POC is a partially ordered protocol which

(20)

3. Transport and Link Layer 7

can also provide partial reliability. POC was later extended to POCv2 [23]. An application speciﬁc use of partial reliability for the transfer of three dimensional modeling data is described in [35]. Other works examining the concepts of partial reliability include [32, 43, 53, 56]. The combination of features that make PRTP unique in relation to earlier work is that PRTP provides deterministic reliability guarantees, can be deployed by only modifying the receiver side, and requires only minor changes to the TCP code and the socket interface.

From the application layer perspective, the desire to make applications more adaptive is a notable trend . The idea that applications should be able to use information about the underlying network conditions to adapt themselves has existed for some time.

Applications can for example adapt by transcoding their data to a different format as network conditions vary, as for example described in [12] for audio and in [11] for video.

Other works examining the use of adaptive applications include [15, 26, 30]

For the more specific application of image transfer over flexible transport protocols, previous work include [37], which focuses on the use of flexible ordering to improve progressive image transfer of GIF images. In [34] the possibility of using a progressively reliable transport protocol for image transfer over wireless links is explored. The Image Transfer Protocol (ITP) [60] runs on top of UDP and is flexible both with regards to order and reliability, and can be used with a suitable coder. It is also possible to involve more protocol layers and create a more specialized solution, for example as done in [39]

which heavily tie the image transfer to a speciﬁc wireless link technology and optimize for that case.

An alternative approach to the use of a ﬂexible transport layer protocol in combination with suitable application support is to adapt only at the application layer. An example of this is the transcoding of images into a more compressed representation as done by transcoding proxies, an approach studied extensively. An early example of a proxy system targeted at a speciﬁc wireless environment (GSM) is described in [42].

Further examples of proxies targeted at wireless networks are [8, 28, 33, 70]. Proxy systems has also been examined from a more general perspective [14, 52, 64]. In addition to the robust JPEG coder, paper III also presents and evaluates resource-efﬁcient methods to perform JPEG transcoding as done in transcoding proxies. Both the proposed transcoding methods, DCT-domain and scan-based, can considerably speed up transcoding compared to general coding tools like compress [36], which is used for transcoding in [64].

3 Transport and Link Layer

3.1 Overview of Area and Thesis Work

Transport protocols can have a considerable inﬂuence on communications performance.

Under ideal circumstances, the transport protocol should ensure that all users get the maximal performance while fairly sharing the available communication resources and

(21)

protecting the network from overloading. Out of the two common transport protocols in the Internet, UDP and TCP, only TCP was designed with consideration to fair sharing and overload protection. The congestion control of TCP has been an important factor that has allowed the Internet to scale up from tens of hosts up to several hundred mil- lion hosts. The use of well functioning congestion control is central for the continued wellbeing of the Internet.

The basic functioning of congestion control is, in simple terms, to react when the network starts to get congested and protect the network by lowering the sending rate.

Incipient congestion will cause buffer overﬂows at routers in the network, and one or more packets will be lost. When a packet loss is detected by the sender, it reduces its sending rate. A TCP sender can detect packet loss either by a timeout or a fast retransmit.

A timeout is triggered when the sender does not receive an acknowledgment (ack) in time. A fast retransmit occurs when the sender has received three acks that acknowledge packets sent after a missing packet. The missing packet is then considered lost and is retransmitted.

There may, however, be cases where these mechanisms do not work ﬂawlessly. One such case is when the underlying layers can have a large variability in delay. This variability may then cause timeouts to occur even though the packet has not been dropped but merely delayed for an unusually long time. These interaction effects may create spurious retransmissions that degrade TCP throughput. An examination of the interaction between the transport and link layers in the GSM environment is reported in Paper IV. This examination was performed using a complete setup of actual GSM hardware in combination with a fading simulator to emulate the physical layer channel conditions.

Another case where the congestion control of TCP can behave inappropriately is when packets are lost, but the cause is not a buffer overﬂow, but some other reason. The major causes for non-congestion related losses is related to wireless links. The physical characteristics of wireless links, with effects such as interference and fading, may lead to corruption which in turn causes packet loss when a subsequent checksum control fails.

Also, the possibility for mobility that is often present when using wireless links might lead to lost packets when nodes are temporarily moving out of coverage. There exists a number of approaches to handle these difﬁculties at various layers in the protocol stack.

Often physical and link layer solutions such as forward error correction (FEC) and link level retransmissions are used to handle the high bit error rates generated by the radio channel. However, there are scenarios where it is not practical or possible for the lower layers to eliminate all bit-errors, typically due to power, complexity or delay constraints.

For many wireless environments the problems surrounding mobility induced losses and intermittent disconnections are relevant and challenging, but these are outside the scope of this thesis. We consider modiﬁcations to the link layer and TCP, which allows TCP to infer the cause of a packet loss and employ a more appropriate congestion behavior for corruption losses. The overall design of checksum-based loss differentiation is discussed in paper V, together with an overview of previous work on loss differentiation. If the loss differentiation is performed on the receiver side, the receiver must be

(22)

3. Transport and Link Layer 9

able to somehow notify the sender of the cause for a loss. The sender, which performs the congestion control, can then apply appropriate behavior for each loss. In paper VI, two different loss notiﬁcation schemes are evaluated from a performance perspective.

Flows that employ loss differentiation and notification typically benefit from increased throughput compared to flows that do not employ it. If these types of flows compete over a shared bottleneck link, the fairness between the flows may be affected by the use of loss differentiation. In paper VII we examine the performance of loss differentiation, and also the fairness between loss differentiation and regular flows.

Heterogeneous networks that include wireless links present additional challenges for the transport layer. We believe that additional ﬂexibility at the transport layer, and increased ﬂow of cross-layer information, can address some of these challenges as illustrated by the work in this thesis.

3.2 Related Work

A large amount of previous work has been made on congestion control modiﬁcations that strive to increase the performance of TCP when used over wireless links. Several survey style publications discuss the various problem and solutions [24, 69, 68].One possible classiﬁcation of these schemes can be in those using split connections [3], TCP-aware local retransmissions (snoop) [6] or loss differentiation at the transport layer [29]. Also, there a some schemes that do not use packet loss as a basis for performing end-to-end congestion control, and are thus less sensitive to corruption losses. Examples of these schemes include TCP Westwood [17] and TCP Real [71].

Transport and link layer interactions in the GSM has been examined before, although not with the same detail of control over the environment. Early work [1] found more negative interaction, whereas our results concur later work in the area [48, 49].

For the case of loss differentiation, a high-level classiﬁcation can be made into those schemes that require support from the infrastructure such as base stations, and those which only need end-host changes. These end-to-end differentiation schemes typically require changes to the sender or the receiver side, although there are schemes that require changes to both sides.

Infrastructure-supported schemes have been proposed on the basis of partial acks provided by the base station [9, 22], a TCP option which is modified by the base station [20], or base station use of queue lengths to infer loss causes [4]. In contrast, end-host loss differentiation located at the sender side tries to process the incoming ack stream to infer the network state. Losses occurring when the network is in a congested state are then classified as congestion losses. If the network, as perceived by the sender, is in a non-congested state, losses are classified as wireless related. A number of ways exist to try to infer the congestion state of the network. Approaches in earlier literature include using a knee point or delay threshold [62], the TCP Vegas style estimation used for TCP Veno [21] and NewReno-LP [50], hidden Markov modeling [47], and machine learning [40]. Common for the sender-based schemes is that they typically have lower

(23)

loss differentiation precision than receiver or infrastructure based differentiation. Their precision also varies with factors such as trafﬁc load, reordering, amount and distribution of cross trafﬁc, buffer sizes, link delays and mechanism parameterization.

End-to-end receiver-based schemes do not have the problem of sensitivity to back- ward channel conditions. Since they are receiver-based, these schemes do, however, require a loss notiﬁcation mechanism to convey the loss differentiation result to the sender.

One earlier work on receiver-based loss differentiation tried to use the inter-arrival times at the receiver to classify the loss causes [10]. This scheme was later improved, and examined along with two other schemes, Spike and ZigZag [18]. An alternate approach is to use checksums to differentiate between loss causes. This has been discussed in earlier work [5, 44], and TCP HACK [7] is a proposal suggesting the use of a TCP option containing a TCP header checksum. Most above differentiation schemes, and others, are further discussed and classiﬁed in paper V.

Other transport layer protocols that consider the possibility of residual bit-errors at the transport layer include UDP-Lite [45], that uses partial checksums and allows the delivery of data with bit-errors to the application. Another example is the new Datagram Congestion Control Protocol (DCCP) [41] that can use a partial checksum option to help differentiate between congestion and corruption losses.

4 Research Issues

4.1 Research Questions

This research covers questions at several levels of abstraction. At the highest level of abstraction the research question that is examined is stated as:

Is it possible to enhance system performance by increased transport layer ﬂexibility and increased ﬂow of cross-layer information to the transport layer?

This general questions has been examined in two specific contexts. The first context relates to flexibility with regards to reliability requirements and increased information exchange between the application and transport layers. The research question related to this specific context is formulated as

Can a web image transfer service beneﬁt from added transport layer ﬂexibility in the form of a partially reliable transport service?

The second context that has been examined relates to link and transport layer interactions. In this context two problem areas were examined, TCP over GSM and the effect of loss differentiation in wireless networks which may have residual bit errors. In this context, two research questions can be formulated:

(24)

4. Research Issues 11

What are the major effects that inﬂuence the performance of TCP over a GSM link?

Can loss differentiation improve the performance of TCP, and what are the fairness im- plications of loss differentiation?

4.2 Research Method

In the computer networks area there are four common approaches to performing protocol evaluation; analytical modeling, simulations, experiments with an emulated network and live network experiments.

Analytical modeling tries to describe the essential behavior of an entity (such as a protocol) with a mathematical expression that given some input parameters provides some metric of interest. When a suitable expression has been derived, it can then easily be used to predict the performance of the entity under a range of conditions. However, in order to create a tractable formula the expression must often be simpliﬁed which intro- duces inaccuracies. This highlights the importance of model veriﬁcation to ensure that the model does not deviate too much from the actual behavior. Nevertheless, modeling can be useful to get approximate behavior that can be mathematically analyzed. The TCP throughput formula proposed by Padhye et al [55] is an example of an analytical model that has become widely used.

The simulation approach also uses abstract representations of the entity under study, but in this case the abstractions are much more detailed and can include all relevant protocol functionality. Also with simulations, there is a need to verify that the abstraction used in the simulation is correct and representative. The simulation approach is very ﬂexible and can handle both small and large network topologies. Simulation allows a large parameter space to be explored and can provide considerable detail in the output.

Commonly used simulation software includes ns2 [51] and opnet [54].

In contrast to analytical modeling and simulation, the emulation approach uses actual protocol implementations running on real hardware. The emulation approach naturally captures the behavior of a protocol implementation, and also takes environmental factors such as possible interaction effects with the operating system, device drivers and communications hardware into account. In the case of emulation it is the behavior of the end-to-end connectivity that is abstracted to some degree. The end-to-end connectivity can be abstracted into one emulated behavior, or a network can be constructed by a number of emulated links. Since each emulated link typically requires one PC, the size of networks with emulated links is restricted for practical reasons. Commonly used emulation tools are NISTnet [16] and dummynet [61].

The live network approach entails performing experiments on a running communications network. This naturally includes all aspects, both at the endpoints and at the network level. However, live experiments are hard to control and repeat. Getting access to live networks to the extent necessary may also be problematic in some instances.

(25)

If the experiments study a behavior that occurs only for speciﬁc network conditions, this requires that these conditions can be introduced in the live network, which may not always be the case.

Most experimental work in this thesis has been performed by using emulation. Paper IV used a fading simulator that emulated the interference caused by multi-path propa- gation of the electromagnetic waves as experienced by a GSM radio receiver. Papers I, II, V, VI and VII used network emulation based on dummynet. To be able to introduce bit errors, and to increase repeatability, we have extended the dummynet network emulator with additional functionality that allows controlled placement of packet losses and bit-errors. The impact of loss generation on network emulation and the design of this dummynet extension is discussed in paper VIII.

With the emphasis of this research being end-point aspects, we believe that the emulation approach provides a good combination of ﬂexibility, representativeness and va- lidity for this work. An limitation of emulation based research is, however, that it is not always possible to perform as many replications as with simulations since emulation experiments run in real time. Consequently, emulation-based experiments sometimes cannot give the same statistical statistical strength as simulations due to the bounds on experimental run time.

5 Contributions

The contributions of this thesis can be related to the research questions posed earlier.

On the most general level the results illustrate that it is indeed possible to improve the performance of communication systems by allowing additional ﬂexibility and inter-layer communication.

Looking a bit more in detail at the first context that was examined, application and transport layer interactions, the results exemplify that improved communications performance is possible with the use of partial reliability. However, for the examined application it is clear that the improved communications performance must be traded of against reduced image fidelity. The performed user perception experiments illustrate the difficulties in arriving in a single quantitative value for how this trade-off should be performed. Qualitatively, several subjects reacted positively to the ability to explicitly control this trade-off.

The second context that was examined in this thesis was transport and link layer interactions. Here, results on GSM link and transport layer interaction reﬂect that while interaction was efﬁcient most of the time, high round-trip time variability could cause negative effects. For the examination of loss differentiation, quantitative results demon- strate that considerable gains can be realized by allowing loss information to be propa- gated to the transport layer. Some of these results are not novel in themselves but rather corroborate and expand on earlier results in the literature using more current live protocol implementations and a wider range of network parameter values. The work in

(26)

6. Summary of Papers 13

this thesis also covers the friendliness of loss differentiation, an area that has not been sufﬁciently examined in previous research.

In addition to the contributions related to the research questions the thesis work has also produced additional benefits to the scientific community. A robust JPEG coder has been developed, capable of transcoding JPEG into a robust data-stream suitable for the transport over a partially reliable transport service. This transcoder can also perform efficient compression transcoding on both sequential and progressive JPEG images.

Valuable insights into experimental research has also been gained during the work, and this has resulted in the development of supporting software. To provide practical means to perform user preference evaluations a distributed client server system has been designed and implemented. Also, the dummynet network emulator has been extended to allow deterministically controlled insertion of bit errors and packet losses.

This dummynet extension has allowed the reported experiments to be performed in a more controlled and repeatable manner, and will also be useful for future research.

6 Summary of Papers

The objective of this thesis is to highlight the possibilities of improving performance in heterogeneous networks by increasing the transport layer flexibility, functionality, and information exchange with other layers. This objective is exemplified by adapting the transport layer behavior in relation to two different layers. Part I of this thesis is concerned with application and transport layer interaction, and this aspect is examined in papers I-III. Part II concerns the interaction between link and transport layer, and this aspect is covered in papers IV-VII. The final paper in the second part, paper VIII, discusses aspects related to the emulation based experimental approach used for this research.

I Application and Transport Layer Flexibility - A User-inﬂuenced Web Browsing System

This paper provides a comprehensive evaluation of the web-browsing application used to evaluate the potential benefits of increased transport layer flexibility in the form of partial reliability. The approach taken in this paper is to first describe the parts of the system, and then proceed to the system level and finally cover user aspects. The partially reliable transport protocol, PRTP, is described and evaluated standalone over a transat- lantic link. The implementation of the robust JPEG coder is described together with a short performance evaluation. A system level examination is reported, enabled by the use of a proxy setup for simple integration of the PRTP protocol and the robust JPEG coder.

The end-users of a web browsing system are individual persons, with varying prior- ities and perceptions. By involving the end-users into the evaluation new insights can be provided. The ﬁnal contribution described in this paper is the user study that was

(27)

performed using the previously described components. The user study highlights the variation between individual users and illustrates the usefulness of extending increased ﬂexibility to cover also the user-level.

II Impact of Partial Reliability on Web Image Transport Performance

This paper expands on the system level experiments included in the previous paper.

This paper contributes with a more detailed analysis of the measurements and explain the special case of late-in-connection losses. This kind of losses can have a considerable impact on the performance of regular TCP connections. PRTP is not sensitive to these kind of losses as long as the reliability requirements can be upheld, which results in a a substantial performance improvement for these cases.

III JPEG Transcoding - Efﬁciency and Robustness Aspects

This paper provides a detailed description of the robust JPEG coder that was used for the experiments described in papers I and II. The robust JPEG coder was speciﬁcally designed with the PRTP protocol in mind and makes it possible to use PRTP’s partially reliable transport service to transfer image data in an efﬁcient manner. Three major extensions were added to standard JPEG in order to allow for partial reliability:

additional restart markers, interleaving and error concealment. These extensions are discussed and the effectiveness of the implemented coder is illustrated. In addition to robustness aspects, this paper also discusses efﬁcient ways of performing compression level transcoding for both sequential and progressive JPEG. For sequential JPEG images, DCT domain transcoding is evaluated and performance measurements presented.

For progressive JPEG, a simple and effective transcoding method called progressive parsing transcoding is presented. Progressive parsing transcoding is based on truncation of the bit stream and allows transcoding to be performed with very little resource usage.

IV GSM / GPRS Testbed for TCP Performance Evaluation

Wireless links are challenging for communications systems due to the inherent physical characteristics of the wireless channel. These challenges are to some extent also present at the transport layer, and the inclusion of a wireless link may have a large impact on the characteristics of an end to end connection. This paper describes a testbed designed to allow the examination of interaction effects between the transport layer and a speciﬁc link layer, that of GSM/GPRS. An important contribution of this paper is the detailed experimental examination of the interaction effects between TCP and the GSM link layer using an actual GSM network dedicated for testing purposes and run by a major Swedish operator. A physical layer emulation of the radio environment made it possible to control and repeat the measurements. The results show that while negative interactions are rare for the examined conditions, the negative interactions that do occur due are primarily

(28)

6. Summary of Papers 15

due to large variations in the round trip times instead of simply being related to poor radio conditions.

V Checksum-based Loss Differentiation

While the previous paper examined transport and link layer layer interactions for a link layer with persistent retransmissions, this paper examines the interactions when the link layer does not persistently retransmit a corrupt frame until it is correctly received. This paper presents the principles behind checksum-based loss differentiation and gives an overview and classiﬁcation of previous work on loss differentiation. The loss differentiation schemes are divided into those that require infrastructure support, end-to-end sender based and end-to-end receiver based. Several schemes in each category are sum- marized and they provide a background to the more detailed description of the proposed checksum-based loss differentiation scheme. A major contribution of this paper lies in the description of the principal issues surrounding checksum-based loss differentiation, such as checksum locality, implementation issues, ﬂow mis-mapping and data integrity.

VI Evaluation of Transport Layer Loss Notiﬁcation

When loss differentiation is performed at the receiver side, as is the case for the checksum- based approach, there will be a need to notify the sender side of the cause for a loss.

Since it normally is the sender that performs the congestion control, this loss notification allows the sender to use appropriate behavior according to the cause for a loss. Loss notification can be done in several ways with different results with regards to ease of employment and performance. The work in this paper extends on the work in the previous paper and describes two different loss notification schemes with different characteristics. Both an implicit 3-dupack mechanism and an explicit TCP-option mechanism are described. Besides presenting a discussion on implementation issues for the two mechanisms, a major contribution of this paper is the experimental evaluation of the performance that can be obtained with the two mechanisms.

VII An Experimental Study on the Performance and Fairness of Loss Differentia- tion for TCP

Compared to paper VI, this paper reports on the performance of loss differentiation with TCP-option notification using additional network parameters. However, the paper also expands the examination to include an alternate method to improve the throughput of TCP over wireless links with residual errors. By using multiple regular TCP connections to transfer data instead of a single one, the total throughput can be increased. This approach has considerable performance benefits over regular, single-connection, TCP when used over links with errors. The approach allows for a different, more application centered, approach to handle problematic links. A final important contribution of this paper is the examination of the fairness aspects of loss differentiation. While loss

(29)

differentiation can provide increased throughput, it also has the potential to unfairly reduce the performance of other ﬂows that do not use loss differentiation. The examination shows that fairness indeed can be an issue, and that fairness considerations are especially relevant for low bandwidth links.

VIII The Impact of Loss Generation on Emulation-based Protocol Evaluation Evaluation of protocol performance can be performed in number of ways. In this thesis, the focus has been on using network emulation to perform protocol performance evaluation. In most cases the evaluations have used randomly generated loss patterns which have then been used to induce losses in a random, but controlled manner. This controlled loss insertion has been made possible by implementing an extension of the well-known dummynet emulator. Controlling the insertion of losses in a deterministic way allows both for getting additional knowledge about speciﬁc protocol behavior by exactly controlling the loss positions, and for the possibility to perform paired tests to gain statistical advantages. This paper provides an overview of the positive effects of using controlled loss generation, and also discusses some implementation details.

7 Conclusions and Outlook

The originally developed transport layer protocols of the Internet, TCP and UDP, have been in successful use during the last two decades. While UDP has been practically constant during these decades, TCP has undergone more evolution to adapt to new requirements and changing environmental conditions. As the network becomes more heterogeneous, however, it in some cases become harder to address the problems only at the transport layer. In this thesis, we have taken a transport layer centered approach, but also involved other layers in order to examine potential benefits of increased cross-layer interaction while upholding the basic principles of layering. Transport layer flexibility was examined in two different contexts using an experimental method based on the use of real implementations in an emulated environment. The focus of the first context is application and transport layer flexibility with regards to reliability. For the case where the network can provide only inadequate bandwidth to a user, the additional flexibility can allow the user to have a more flexible trade-off between waiting time and image quality when browsing the web. To this end, we have developed a suitable JPEG-based image coder to examine the system level effects of partial reliability. The experimental results show that gains can be achieved, and that the gain varies between different compositions of web-pages. A user study was performed highlighting the individual differences in user preferences and the relevance of being able to perform the trade-off on a per user basis.

Also examined in this thesis is transport and link layer ﬂexibility. The case of residual bit-errors on the wireless link was studied and a checksum-based loss differentiation mechanism was developed. Together with a loss notiﬁcation mechanism this allows for

(30)

REFERENCES 17

differentiation between loss causes and the use of more appropriate behavior for link- induced losses. The performance was evaluated using two different loss notification schemes, one implicit and one explicit. While the implicit 3-dupack scheme improved performance in several cases, it did not have the overall effectiveness of the explicit TCP- option scheme. In addition to performance, fairness was also examined. The fairness results are interesting since they show that the improved performance of loss differentiation may indeed degrade the performance of competing non-differentiating flows. This is especially relevant for low bandwidth connections, where the improved throughput for loss-differentiating flows was matched by decreased throughput for non-loss differentiating flows. The same mechanisms that lead to improved throughput effects, may thus in other cases lead only to a decrease in fairness. This result can be assumed to be valid for other similar loss differentiating techniques, making it interesting for further study. The effects of GSM link layer retransmissions on TCP performance was also examined. The results of this study concluded that when negative interactions do occur, they are coupled to large variability in link layer round-trip times rather than simply to poor radio conditions.

While the mechanisms studied in this thesis were studied in the context of TCP, they are more generally applicable to other transport layer protocols. We notice that the mechanisms examined in this thesis have been included in recent IETF protocols.

The concept of partial reliability, to which the ﬁrst part of this thesis relates, has been standardized as an extension to the SCTP protocol [66]. Checksum-based loss differentiation mechanisms, as studied in the second part, are included in the proposed DCCP protocol [41]. Studying the potential of these mechanism in the setting of these new protocols is a possible avenue for further research. Additional future work includes further enhancement of the emulation-based experimental platform developed and used in this research. In this context a cooperation within the EU network of excellence NEWCOM has already been initiated to integrate a scenario generator for ad-hoc networks research.

References

[1] T. Alanko, M. Kojo, H. Laamanen, M. Liljeberg, M. Moilanen, and K. Raatikainen.

Measured performance of data transmission over cellular telephone networks.

ACM Computer Communication Review, 24(5), 1994.

[2] P. D. Amer, C.Chassot, T. Connolly, M. Diaz, and P. Conrad. Partial order trans- port service for multimedia and other applications. IEEE/ACM Transactions on Networking, 2(5):440–456, October 1994.

[3] A. Bakre and B. R. Badrinath. I-TCP: Indirect TCP for mobile hosts. 15th Inter- national Conference on Distributed Computing Systems, May 1995.

(31)

[4] H. Balakrishnan and R. Katz. Explicit loss notiﬁcation and wireless web perfor- mance. In Proceedings Globecom Internet Mini-Conference, Sydney, Australia, November 1998.

[5] H. Balakrishnan, V. N. Padmanabhan, S. Seshan, and R. H. Katz. A comparison of mechanisms for improving TCP performance over wireless links. IEEE/ACM Transactions on Networking, 5(6):756–769, December 1997.

[6] H. Balakrishnan, S. Seshan, E. Amir, and R. H. Katz. Improving TCP/IP perfor- mance over wireless networks. In proc. 1st ACM Int’l Conf. on Mobile Computing and Networking (Mobicom), November 1995.

[7] R. K. Balan, B. P. Lee, K. R. R. Kumar, L. Jacob, W. K. G. Seah, and A. L. Ananda.

TCP HACK: A mechanism to improve performance over lossy links. Computer Networks, 39(4):347–361, July 2002.

[8] H. Bharadvaj, A. Joshi, and S. Auephanwiriyakul. An active transcoding proxy to support mobile web access. Proceedings of the 17th IEEE Symposium on Reliable Distributed Systems, 1998.

[9] S. Biaz, M. Mehta, S. West, and N. H. Vaidya. TCP over wireless networks using multiple acknowledgements. Technical Report 97-001, Computer Science, Texas A&M University, January 1997.

[10] S. Biaz and N. Vaidya. Discriminating congestion losses from wireless losses using inter-arrival times at the receiver. IEEE Symposium ASSET’99, Richardson, TX, USA, March 1999.

[11] J.-C. Bolot and T. Turletti. Adaptive error control for packet video in the Internet.

In Proceedings CIP ’96, Lausanne, September 1996.

[12] J.-C. Bolot and A. Vega-Garcia. Control mechanisms for packet audio in the Inter- net. In Proceedings IEEE INFOCOM’96, San Fransisco, CA, March 1996.

[13] T. Boutell. RFC 2083: Png (portable network graphics) speciﬁcation, March 1997.

[14] C. Brooks, M. S. Mazer, S. Meeks, and J. Miller. Application-speciﬁc proxy servers as http stream transducers. Proc. 4th Intl WWW Conference, Boston MA, December 1995.

[15] A. Campbell and G. Coulson. A QoS adaptive transport system: design, imple- mentation and experience. In Proceedings ACM Multimedia 96, pages 117–127, 1996.

[16] M. Carson and D. Santay. NIST Net: A linux-based network emulation tool. ACM SIGCOMM Computer Communication Review, 33(3):111–126, 2003.

(32)

REFERENCES 19

[17] C. Casetti, M. Gerla, S. Mascolo, M. Sansadidi, and R. Wang. TCP Westwood:

End-to-end congestion control for wired/wireless networks. Wireless Networks, 8(5):467–479, September 2002.

[18] S. Cen, P. Cosman, and G. Voelker. End-to-end differentiation of congestion and wireless losses. Proc. Multimedia Computing and Networking (MMCN2002), San Jose, CA, pages 1–15, January 2002.

[19] V. Cerf and R. Kahn. A protocol for packet network interconnection. IEEE Trans- actions on Communications, 22:637–648, March 1974.

[20] W.-P. Chen, Y.-C. Hsiao, J. C. Hou, Y. Ge, and M. P. Fitz. Syndrome: a light-weight approach to improving TCP performance in mobile wireless networks. Wireless Communications and Mobile Computing, (2):37–57, 2002.

[21] F. Chengpeng. TCP Veno: End-to-end congestion control over heterogeneous net- works. Ph. D thesis, Chinese University of Hong Kong, July 2001.

[22] J. A. Cobb and P. Agrawal. Congestion or corruption? A strategy for efﬁcient wireless TCP sessions. IEEE Symposium on Computers and Communications, Alexandria, Egypt, pages 262–268, 1995.

[23] P. T. Conrad, E. Golden, P. D. Amer, and R. Marasli. A multimedia document retrieval system using partially-ordered/partially-reliable transport service. Pro- ceedings of Multimedia Computing and Networking, January 1996.

[24] S. Dawkins, G. Montenegro, M. Kojo, V. Magret, and N. Vaidya. RFC 3155:

End-to-end performance implications of links with errors, August 2001.

[25] B. J. Dempsey. Retransmission-based error control for continuous media. PhD thesis, University of Virginia, 1994.

[26] C. Diot, C. Huitema, and T. Turletti. Multimedia applications should be adaptive.

In Proceedings HPCS’95, Mystic (CN), August 1995.

[27] M. B. (ed.). JPEG 2000 Final Committee Draft. ISO/IEC FCD15444-1, March 2000.

[28] A. Fox and E. A. Brewer. Reducing WWW latency and bandwidth requirements by real-time distillation. Proc. Fifth International World Wide Web Conference, Paris, France, May 1996.

[29] J. Garcia and A. Brunstrom. Checksum-based loss differentiation. Proceedings 4th IEEE Conference on Mobile and Wireless Communications Networks (MWCN 2002), Stockholm, Sweden, September 2002.

[30] J. Gecsei. Adaption in distributed multimedia systems. IEEE Multimedia, 4(2):58–

66, Apr.–June 1997.

(33)

[31] Graphics Interchange Format, version 89a. Technical Report, Compuserve Incor- porated, Columbus, Ohio, July 1989.

[32] F. Gong and G. Parulkar. An application-oriented error control scheme for high- speed networks. IEEE/ACM Transactions on Networking, 4(5):669–683, October 1996.

[33] R. Han, P. Bhagwat, R. LaMaire, T. Mummert, V. Perret, and J. Rubas. Dynamic adaptation in an image transcoding proxy for mobile www browsing. IEEE Per- sonal Communication, 5(6):8–17, December 1998.

[34] R. Han and D. Messerschmitt. A progressively reliable transport protocol for inter- active wireless multimedia. Multimedia Systems, Springer-Verlag, 7(2):141–156, 1999.

[35] A. Harris and R. Kravets. The design of a transport protocol for on-demand graph- ical rendering. Proc 12th International Workshop on Network and Operating Sys- tems Support for Digital Audio and Video (NOSSDAV 2002), Miami Beach, FL, May 2002.

[36] Imagemagick suite. http://www.imagemagick.org/, April 2005.

[37] S. Iren. Network-conscious Image Compression. PhD thesis, University of Delaware, 1999.

[38] ITU-T. Recommendation T.81 - digital compression and coding of continuous-tone still images. Geneva, Switzerland, September 1992.

[39] L. Jianhua, M. L. Liou, K. B. Letaief, and J. C.-I. Chuang. Mobile image transmission using combined source and channel coding with low complexity concealment.

Signal Processing: Image Communication, 12(2):87–104, April 1998.

[40] I. E. Khayat, P. Geurts, and G. Leduc. Improving TCP in wireless networks with an adaptive machine-learnt classiﬁer of packet loss causes. In 4th International IFIP-TC6 Networking Conference (NETWORKING 2005), Waterloo, Canada, May 2005.

[41] E. Kohler, M. Handley, and S. Floyd. Datagram congestion control protocol (DCCP). draft-ietf-dccp-spec-08.txt, Work in progress, October 2004.

[42] M. Kojo, K. Raatikainen, and T. Alanko. Connecting mobile workstations to the In- ternet over a digital cellular telephone network. Proceedings Workshop on Mobile and Wireless Information Systems (Mobidata), Rutgers University, NJ, November 1994.

[43] R. Kravets, K. L. Calvert, P. Krishnan, and K. Schwan. Adaptive variation of relia- bility. Proc Seventh IFIP Conference on High Performance Networking (HPN’97), White Plains, April 1997.

(34)

REFERENCES 21

[44] R. Krishnan, M. Allman, C. Partridge, and J. P. Sterbenz. Explicit transport error notiﬁcation for error-prone wireless and satellite networks. BBN Technical Report No. 8333, BBN Technologies, March 2002.

[45] L.-A. Larzon, M. Degermark, S. Pink, L.-E. Jonsson, and G. Fairhurst. RFC 3828:

The lightweight user datagram protocol (udp-lite), July 2004.

[46] B. Leiner, R. Cole, J. Postel, and D. Mills. The DARPA internet protocol suite.

IEEE Communications Magazine, 23:29–34, March 1985.

[47] J. Liu, I. Matta, and M. Crovella. End-to-end inference of loss nature in a hybrid wired/wireless environment. CS Dept Technical report 2002-008, Boston Univer- sity, March 2002.

[48] R. Ludwig and R. H. Katz. The Eifel algorithm: Making TCP robust against spurious retransmissions. ACM Computer Communication Review, 30(1):30–37, January 2000.

[49] R. Ludwig, B. Rathonyi, A. Konrad, K. Oden, and A. Joseph. Multi-layer tracing of TCP over a reliable wireless link. ACM SIGMETRICS, 1999.

[50] F. Martignon and L. Fratta. Loss differentiation schemes for tcp over wireless networks. In Workshop on Quality of Service in Multiservice IP Networks (QoS-IP 2005), Catania, Italy, February 2005.

[51] S. McCanne and S. Floyd. The Network Simulator - ns-2.

http://www.isi.edu/nsnam/ns/, April 2005.

[52] R. Mohan, J. Smith, and C.-S. Li. Adapting multimedia internet content for uni- versal access. IEEE Transactions on Multimedia, 1(1):104–114, March 1999.

[53] B. Mukherjee and T. Brecht. Time-lined TCP for the TCP-friendly delivery of streaming media. In Proceedings International Conference on Network Protocols (ICNP), Osaka Japan, pages 165–176, November 2000.

[54] OPNET Modeler. http://www.opnet.com/, April 2005.

[55] J. Padhye, V. Firoiu, D. Towsley, and J. Krusoe. Modeling TCP throughput: A sim- ple model and its empirical validation. Proceedings of the ACM SIGCOMM ’98 conference on Applications, technologies, architectures, and protocols for com- puter communication, pages 303–314, August 1998.

[56] M. Piecuch, K. French, G. Oprica, and M. Claypool. A selective retransmission protocol for multimedia on the Internet. In Proceedings SPIE Multimedia Systems and Applications, pages 79–90, November 2000.

[57] J. Postel. RFC 768: User Datagram Protocol, Aug. 1980.

(35)

[58] J. Postel. RFC 791: Internet Protocol, Sept. 1981.

[59] J. Postel. RFC 793: Transmission control protocol, Sept. 1981.

[60] S. Raman, H. Balakrishnan, and M. Srinivasan. ITP: An image transport protocol for the Internet. In Proceedings Intl. Conference on Network Protocols, November 2000.

[61] L. Rizzo. Dummynet: A simple approach to the evaluation of network protocols.

ACM Computer Communication Review, 27(1):31–41, January 1997.

[62] N. K. G. Samaraweera. Non-congestion packet loss detection for TCP error recov- ery using wireless links. IEE Proc. Comm., 146(4):222–230, 1999.

[63] H. Schulzrinne, S. Casner, R. Frederick, and V. Jacobson. RFC 3550: RTP: A transport protocol for real-time applications, July 2003.

[64] A. Singh, A. Trivedi, K. Ramamrithan, and P. Shenoy. PTC: Proxies that transcode and cache in heterogeneous web client environments. World Wide Web: Internet and Web Information Systems, 7(1):7–28, March 2004.

[65] R. Stewart. RFC 2960: Stream Control Transmission Protocol, October 2000.

[66] R. Stewart, M. Ramalho, Q. Xie, M. Tuexen, and P. Conrad. RFC 3758 - Stream control transmission protocol (SCTP) partial reliability extension, May 2004.

[67] A. S. Tanenbaum. Computer Networks. Prentice-Hall International, Inc., 2003.

[68] A. Wennstr¨om, S. Alfredsson, and A. Brunstrom. TCP over wireless networks.

Karlstad University Studies 2004:21, Karlstad University, 2004.

[69] G. Xylomenos, G. Polyzos, P. Mahonen, and M. Saaranen. Tcp performance issues over wireless links. IEEE Communications Magazine, 39(4):52–58, April 2001.

[70] B. Zenel and D. Duchamp. A general purpose proxy ﬁltering mechanism applied to the mobile environment. Mobicom’97, October 1997.

[71] C. Zhang and V. Tsaoussidis. TCP Real: Improving real-time capabilities of TCP over heterogeneous networks. Proceedings of the 11th IEEE/ACM NOSSDAV, 2001.

[72] H. Zimmerman. OSI reference model - The ISO model of architecture for open systems interconnection. IEEE Transactions on Communications, 28(4):425–432, April 1980.

(36)

)MPROVING (ETEROGENEOUS

4HE GENCE PABILITIES ARRAY HIGHER LAYER NETWORK END TO

4HE ABILITY IS PERIMENTAL INDICATE

4HE CONSEQUENCE RESIDUAL DEGRADATION

NISMS PERFORMANCE CONSIDERED

+ARLSTAD

)33.

)3".

Improving Performance in Heterogeneous Networks: A Transport Layer Centered Approach

*OHAN

)MPROVING (ETEROGENEOUS

*OHAN

)MPROVING (ETEROGENEOUS

Improving Performance in Heterogeneous

Networks: A Transport Layer Centered Approach

Abstract

Acknowledgments

List of Appended Papers

Comments on my participation

Other Work

Contents

Introductory Summary

1 Introduction

2 Application and Transport Layer

2.1 Overview of Area and Thesis Work

2.2 Related Work

3 Transport and Link Layer

3.1 Overview of Area and Thesis Work

3.2 Related Work

4 Research Issues

4.1 Research Questions

4.2 Research Method

5 Contributions

6 Summary of Papers

7 Conclusions and Outlook

References

)MPROVING (ETEROGENEOUS