Modeling and Performance Evaluation of TCP over Last-Mile Wireless Networks

Transcription

1 Modeling and Performance Evaluation of TCP over Last-Mile Wireless Networks Shiva Raj Pokhrel Faculty of Science Engineering and Technology Swinburne University of Technology A thesis submitted for the degree of Doctor of Philosophy 2017

2 My thesis is dedicated to my love Su Surakshya, my cutey Susie Susikshya, my lovely parents Pad Padam and Kali Kalika, and my all relatives and friends, whose love and support made this possible.

3 Acknowledgements This thesis was completed with the support of many people, to whom I would like to express my sincere thanks. First and foremost, I would like to give special thanks to my supervisors, Prof. Hai Le Vu, Dr. Antonio Cricenti and Dr. Manoj Panda, for their great supervision during the past three and a half years. All the work described in this thesis was finished under their guidance. I would like to give special thanks to Prof. Michel Mandjes, University of Amsterdam (UvA) for his guidance and generous support throughout my time of PhD. All of them supported me with valuable research ideas, constructive comments, efficient research methodology as well as constant encouragement and endless patience. I have gained much knowledge in doing research from their instruction, which will be very useful for my future career. I greatly appreciate the generous support of Prof. Grenville Armitage and A/Prof. Philip Branch for all students in the School of Software and Electrical Engineering at Faculty of Science, Engineering and Technology (FSET). Also thanks to FSET, where I spent all my four years of PhD candidature and made friends with many friendly, helpful and knowledgeable researchers and students, with whom discussion has been very interesting and useful for my own knowledge. I am grateful to all my friends at FSET for all the emotional support, entertainment, and caring they provided. Thanks for treating me like a family member. This thesis would not have been possible without the strong support, encouragement and love of my spouse, parents and my daughter. They have always been there for me to encourage on throughout my time of PhD candidature.

4 Abstract Wireless networks are characterized by a shared transmission media with varying network congestion and channel conditions, which presents a challenge and a formidable barrier to the reliable transmission of digital data. In recent years, these problems have been dealt with the gradual advancements in networking algorithms and physical layer transmission technologies. However, a robust and reliable high-speed data communication over last-mile wireless networks is still challenging, which is currently impending progress. The most commonly used Internet Protocol (IP) based standard over the widely deployed broadband last-mile wireless networks is the Transmission Control Protocol (TCP). The TCP/IP based communication is largely restricted to a single path per connection; yet, multiple paths often exist between the communicating devices. The simultaneous use of these multiple paths for a TCP/IP connection would improve resource usage within the network and, thus, improve user experience through higher throughput and improved resilience to network failure. Multipath TCP, recently standardized by IETF, provides the ability to concurrently use multiple paths between peers. The objective of this thesis is to investigate and exploit the dynamics of multipath TCP and its congestion control over broadband wireless networks. Common examples of broadband wireless networks deployed today include WiFi networks that operate in the unlicensed frequency bands and broadband Cellular networks such as 2G, 3G and LTE that operate in the licensed bands. When TCP is used as the transport protocol over these networks, it runs into various problems which are not commonly found in wired networks; some of these are: 1. misinterpretation of higher channel loss rates as congestion; 2. higher delay;

5 3. large delay variations; and 4. asymmetric path capacities and characteristics. Huge efforts have been made over the past decade to find ways by which the wireless channels can be made more reliable and highly performing. A lot of work has gone into improving the wireless physical layer with more robust transmission and powerful error correction and coding schemes, but equally important have been the advances in the Medium Access Control (MAC) and Transport Layers. Some of this effort has also gone into analytic modeling for modifying the congestion dynamics in order to overcome wireless problems. The analytic modeling approach to enhance the performance of TCP over last-mile wireless networks is the research direction taken up in this thesis. This thesis is primarily concerned with 1. developing the analytic models of regular (single-path) and multipath TCP over the last mile Cellular and WiFi networks; 2. investigating the fairness of bandwidth sharing among users; 3. identifying the root causes of unfairness, if any, and obtaining insights as to how fairness could be reinstated. We develop accurate models and propose efficient solutions for throughput fairness. Firstly, we perform an analysis for a WiFi network acting as a last-mile Internet access with multiple long-lived TCP connections on both the up and down links. This model considers the joint impact of buffer losses at the access point, contention at the medium access control layer, and packet losses due to the wireless channel being erroneous. We show that the model accurately quantifies the probability of an arbitrary TCP packet being discarded, and the total throughput obtained on the up and down links. We find that both the wireless channel errors and buffer overflows lead to throughput unfairness, but that they do so in the opposite direction on the up and down links respectively. We demonstrate that this insight can be exploited so as to significantly mitigate the throughput unfairness. Next, we find that the behavior of multipath TCP over last-mile wireless networks is different from what was expected. We propose several possible improvements by controlling the parameters of WiFi and Cellular 5

6 networks. We develop a comprehensive mathematical model for multiple long-lived multipath connections downloading content from a remote server in the Internet using parallel WiFi and Cellular paths. To the best of our knowledge, it is the first analytical model developed in the literature that captures the coupling between the paths through heterogeneous wireless networks where the coupling arises due to the multipath TCP coupled congestion control protocol. This model also takes into account the impact of the shared nature of the wireless medium and the finite access point (AP) buffer in the last-mile WiFi. Furthermore, we discover a new type of throughput unfairness among the competing regular and multipath TCP connections going through the same AP with a droptail buffer; the regular TCP connections essentially steal almost all the WiFi bandwidth away from the multipath TCP connections. Our steady state analytical model provides the appropriate value of the equal admission probability into the AP buffer to alleviate the throughput unfairness. Finally, we perform an integrated fluid and packet-level analytic modeling to study the transient dynamics of a system of coexisting regular and multipath TCP users over the WiFi and Cellular networks. However, the throughput unfairness between regular and multipath TCP as mentioned earlier is detrimental. In order to fix the temporal fairness issue, we develop an intelligent realtime loss management algorithm that is based on a closed-form expression derived from the integrated model, which solves fairness dynamically. Moreover, we identify the design properties of the algorithm and guarantee the existence, uniqueness and stability of the system equilibrium. 6

7 Declarations This is to certify that 1. the thesis comprises only my original work, 2. due acknowledgement has been made in the text to all other material used, 3. the thesis is less than 100,000 words in length, exclusive of table, figures, bibliographies, appendices and footnotes. Signature: Sthadmk! Date: 14/07/2017 7

8 Contents 1 Introduction Thesis Statement Regular TCP Performance over WiFi Contributions Performance Analysis of Multipath TCP Contributions Co-existence of Regular and Multipath TCP Contributions Organization of the Thesis Publications Included in the Thesis Not included in the Thesis Background and Literature Review Last-Mile Wireless Access Networks Cellular Access Technologies Cellular Architecture WiFi Architecture WiFi Standards Physical Layer Medium Access Control Layer Cellular and WiFi Data Networks TCP Protocols Regular TCP and Single Path Congestion Control Multipath Congestion Control Connection and Subflow Establishment Overview of Multipath TCP Algorithms Network Queue Management i

9 Active Queue Management (AQM) Emerging Concepts TCP Modelling Techniques Fairness and Utility Functions Optimization and Control Theoretic Analysis LaSalle s Invariance Principle Nash Equilibrium State of the Art of TCP Models Modelling TCP over Wireless Conclusions TCP Performance over WiFi Introduction Related Works Network Scenario and Settings Analytical Model Impact of Wireless Channel Errors Tracking the number of active STAs Failure and discard analysis Computation of unrestricted throughput Saturation contention analysis Computing h Actual throughput accounting for the window limit With Buffer and Wireless Channel Errors Results Simulation Setting Model Validation Impact of bottleneck buffer overflow without channel errors Impact of wireless channel errors Joint impact of wireless errors and buffer losses Causes of Throughput Unfairness Design of Solution to Achieve Fairness Fairness with Buffer Sizing Fairness with Admission Control Conclusions ii

10 4 Modeling Multipath TCP Introduction Related Works Network Settings and Mechanism Modeling Assumptions and Ideas Assumptions Main Ideas Modeling of the Homogeneous Case Modeling of the Cellular Path Cellular Window Analysis Computation of Throughput and Probability of Buffer Loss over the Cellular Path Packet Discards due to Wireless Errors Modeling of the WiFi Path RTT Computation for the WiFi Path WiFi Window Analysis Computation of Sending Rate over the WiFi Path AP Buffer Loss Analysis Tracking the Number of active WiFi Users Computation of Aggregate WiFi Throughput Saturation Contention Analysis Impact of Wireless Channel Errors The Fixed Point Iteration Model Validation Modeling and Ensuring Fairness the Mixed Case Reasons for the Observed Unfairness Design of Solution to Reduce Unfairness Loss Equalization Computation of the Split of WiFi Throughput The Nested Fixed Point Iteration Implementation of the Solution and Results Performance of the Solution Approaches Summary of NS-2 Simulation Processes Conclusion iii

11 5 Co-existence of Regular and Multipath TCP Introduction Related Works Network Scenario and Assumptions An Integrated Fluid/Packet Analytical Model Fluid Model for Regular TCP Reno Fluid Model for Multipath TCP LIA Queue Dynamics WiFi Throughput and Discard Probability Cellular throughput and Discard probability Loss Probability for Fairness Model Validation Adaptive Loss Management Algorithm Practical Considerations in the Algorithm Design Loss Dynamics on the Cellular Path Stability Analysis of the ALM Algorithm Performance of the ALM Algorithm Accuracy of the Proposed Algorithm TCP Friendliness and Convergence Overview of Simulation Conclusion Conclusions, Discussions and Future Recommendations Contributions Discussions and Future Recommendations Final Remarks A Derivation and Proofs in Chapter A.1 Saturation Contention Analysis A.2 Existence and Convergence of Fixed Point Solution B Derivation and Proofs in Chapter B.1 Saturation Contention Analysis B.2 Proof of Theorems B.2.1 Proof of Theorem (Existence and Uniqueness) B.2.2 Proof of Theorem (Convergence and Stability) iv

12 Bibliography 171 v

13 List of Figures 1.1 Structure of the Thesis An infrastructure WLAN consisting of an AP, uploading STAs and downloading STAs. The STAs communicate with servers in the Internet through the AP using TCP connections Multipath TCP users are downloading data from a remote server in the Internet using their WiFi and Cellular interfaces. The WiFi Access Point (AP) is shared by all multipath TCP connections. The Cellular paths of the multipath connections have dedicated buffer and wireless channel at the Base Station (BS) Evolution of cellular mobile standard from 2G to 5G GSM architecture Ad-hoc Mode and Infrastructure Mode of WiFi Architecture Evolution of MAC channel activity in IEEE DCF Three way handshake of TCP Handshake of the initial Multipath TCP subflow Handshake of additional Multipath TCP subflows An infrastructure WLAN consisting of an AP, N u uploading STAs and N d downloading STAs. The STAs communicate with servers in the Internet through the AP using long-lived TCP connections. The servers are connected to the AP by high-speed wired links with negligible delay Saturated AP with wireless channel error probability, p w and 5 upload and 5 download long lived TCP connections as observed in NS-2 simulations Analytical block diagram. The block shown by a dotted rectangle exists only for the case of finite AP buffer. The block for computation of h is different for the infinite and finite buffer cases. All other blocks are identical for the infinite and finite AP buffer vi

14 3.4 Evolution of network activity in the WLAN. The time instants immediately after the completion of a service have been marked by arrows. The time between two consecutive arrows is referred to as a cycle. A cycle terminates with either a successful transmission or a discard State transition diagram of {(D k, U k )} Throughput unfairness due to buffer overflows with N u = 5, N d = 5, W max = 20 and bottleneck AP buffer (B ap ) Probability of transmission failure, γ with N u = 5 and N d = 5 and wireless channel error, p w Discard probability, p l with N u = 5 and N d = 5 and wireless channel error, p w Throughput unfairness due to wireless channel error, p w, with N u = 5 and N d = 5 and without buffer overflows at bottleneck AP Analytical and Simulation results depicting Throughput unfairness with N u = 5 and N d = 5 and joint impact of wireless error, p w = 0.2 and AP buffer size, B ap Evolution of the congestion windows of upload and download connections with W max = 45 and p w = 0.3 as observed in NS Evolution of the congestion windows of upload and download connections with W max = 45 and p w = 0.3 as observed in NS Scaling of throughput unfairness with receiver window, W max, with infinite bottleneck AP buffer as observed in NS-2 simulations Translation of throughput unfairness with smaller retransmission, K, with infinite AP buffer as observed in NS-2 simulations Buffer sizing with p w = 0.2 and N u = N d = Buffer sizing with p w = 0.3 and N u = N d = Buffer sizing with p w = 0.3 and N u = N d = Admission control with p w = 0.2 and N u = N d = Admission control with p w = 0.3 and N u = N d = Admission control with p w = 0.4 and N u = N d = vii

15 4.1 M r regular and M m multipath TCP users are downloading data from a remote server in the Internet using their WiFi and Cellular interfaces. The WiFi Access Point (AP) is shared by the M r regular and the M m multipath TCP connections. The Cellular paths of the multipath connections have dedicated buffer and wireless channel at the Base Station (BS) Evolution of short-term individual user throughputs with four regular TCP and four multipath TCP connections going through the AP. The throughputs obtained with a droptail buffer (the topmost and bottommost trajectories) exhibit severe unfairness as regular and multipath TCP connections experience unequal loss rates at the AP buffer. The observed throughputs with the wired case is comparable to that provided by our solution (the trajectories in the middle) Evolution of the AP buffer occupancy as observed in an NS-2 simulation run with five multipath TCP connections sharing an AP buffer of B ap =100 packets Evolution of a multipath connection s congestion windows pertaining to the Cellular path and WiFi path (rounded off to their nearest integer values), denoted by X c and X w, respectively, as observed in an NS-2 simulation run with five multipath TCP users, a BS buffer of B bs =100 packets, AP buffer of B ap =100 packets and channel error probability of 30%; other settings are as in Tab Block diagram of the analytical model Evolution of a multipath connection s congestion windows pertaining to the Cellular path and WiFi path (rounded off to their nearest integer values), denoted by X c and X w, respectively, as observed in an NS-2 simulation run with five multipath TCP users, a BS buffer of B bs =100 packets and an AP buffer of B ap =100 packets; other settings are as in Tab Evolution of states in the shared WiFi channel. The time instants immediately after the completion of a service have been marked by arrows. The time between two consecutive arrows is a so-called cycle State transistion diagram of the process {U k } viii

16 4.9 Throughputs obtained by a multipath TCP user over the Cellular and WiFi paths as a function of the number of users, M m. Notice that θ m obtained from simulations with a droptail buffer at the AP is very close to that obtained with admission control at the AP using admission probability 1 p m, where p m is the loss probability computed using our analytical model Throughputs obtained by a multipath TCP user over the Cellular path as a function of the number of users, M m Buffer loss probabilities experienced by a (long-lived) regular or multipath TCP user at the AP and the BS as a function of the number of users, M r, or M m. Note that these results correspond to only one type of users, i.e. either regular or multipath TCP, but not mixed Discard probability with increasing wireless channel error experienced by five competing multipath TCP users in the Cellular and WiFi path with B ap = 100 and B bs = Throughputs with increasing wireless channel error experienced by five competing multipath TCP users in the Cellular and WiFi path with B ap = 100 and B bs = Block diagram of the model for the mixed case Variation of individual user throughputs with (equal) number of regular and multipath TCP connections for different solution approaches. Both Cellular and WiFi throughputs from analysis and simulations are shown for comparison. Note that the actual buffer requirement with probabilistic admission control Variant II is larger than B ap = 100 packets.probabilistic admission control Variant I with B ap = 100 packets Variation of individual user throughputs with (equal) number of regular and multipath TCP connections for different solution approaches. Both Cellular and WiFi throughputs from analysis and simulations are shown for comparison. Note that the actual buffer requirement with probabilistic admission control Variant II is larger than B ap = 100 packets.probabilistic admission control Variant II with B ap = 100 packets ix

17 4.17 Variation of individual user throughputs with (equal) number of regular and multipath TCP connections for different solution approaches. Both Cellular and WiFi throughputs from analysis and simulations are shown for comparison. Note that the actual buffer requirement with probabilistic admission control Variant II is larger than B ap = 100 packets. Buffer splitting with B ap =100 packets Variation of the difference between individual user throughputs as the number of users of one kind changes and the other kind is kept fixed Variation of individual user throughputs w.r.t. the size of the AP or BS buffer (in packets) Variation of buffer loss probabilities w.r.t. the size of the AP or BS buffer (in packets) Unfairness between regular and multipath TCP flows (four of each type) with droptail buffering at the WiFi access point buffer of capacity 100 packets as observed in NS-2 simulations. θ r (t) and θ m (t) denote the time-varying throughputs (in packets per second) obtained by a regular and an multipath TCP connection, and p r (t) and p m (t) are the corresponding loss probabilities There are n r (t) regular and n m (t) multipath TCP users at time t downloading data from a remote server in the Internet using their WiFi and Cellular interfaces. The WiFi Access Point (AP) is shared by the regular TCP connections and the WiFi paths of the multipath TCP connections. The Cellular path of each multipath TCP connection is allocated with a dedicated buffer and a dedicated wireless channel at the Cellular Base Station (BS) Evolution of the AP buffer occupancy as observed in an NS-2 simulation run with five multipath TCP connections sharing an AP buffer of B ap =100 packets in the presence of wireless channel error probability, p w = 20% Interrelationships between the analytical components and ALM algorithm. The ALM algorithm we propose in this chapter replaces the loss analysis block and ensure interprotocol throughput fairness Typical state transistions of the Markov chain {(R k, M k )}. Here, p d (t) is the discard probability after maximum number of retransmissions. Note that only a small part of the actual chain is shown here x

18 5.6 Individual WiFi throughput in packets/sec of four TCP and four MPTC in the presence of channel errors Individual Cellular throughput in packets/sec of four multipath TCP in the presence of channel errors The accuracy of the approximation η ˆη: Comparison of numerical results with η and ˆη obtained using the model where the number of flows {n r (t), n m (t)} vary as (4, 4) (6, 4) (4, 4) (4, 6) The accuracy of the model with ˆη and responsiveness of ALM algorithm: Comparison of simulations and numerical results with ˆη where the number of flows {n r (t), n m (t)} vary as (4, 4) (6, 4) (4, 4) Magnification of 5.9: Observe the TCP friendliness of ALM algorithm where the throughput of an multipath TCP connection over the WiFi path is approximately 16% lower than that of a regular TCP connection Numerical results showing the impact of ˆη where the number of flows {n r (t), n m (t)} vary as (4, 4) (6, 4) (4, 4). Compare the TCP friendliness and speed of convergence of the ALM algorithm with two different magnitude of η Numerical results showing the impact of the gain parameter m where the number of flows {n r (t), n m (t)} vary as (4, 4) (6, 4) (4, 4). m(t) chosen with gradient is compared with m(t) = 100t. Observe the TCP unfriendliness and throughput oscillations with m(t) = 100t Numerical and simulation results showing the impact of the wrong choice for the gain parameter as m(t) = 100t where the number of flows {n r (t), n m (t)} vary as (6, 4) (4, 4). Observe the TCP unfriendliness and throughput oscillations with m(t) = 100t in both simulations and numerical results Variation in queue length with and without channel errors as observed in NS-2 simulation with 5 uploading and 5 downloading stations and a bottleneck AP buffer of B ap = 60 packets Variation in queue length with different channel errors as observed in NS-2 simulation with 5 uploading and 5 downloading stations and a large bottleneck AP buffer xi

19 6.3 Variation in throughput over time for long-lived TCP flows as observed in simulations with a large AP buffer of B ap = packets. Plots with black and red markers correspond to the throughputs obtained by 3 uploading and 3 downloading connections, respectively, with a wireless channel error probability of p w = 0.3. Throughput obtained by different STAs are marked with different symbols Variation in throughput with the size of short-lived flows as observed in simulations with a large AP buffer of B ap = packets. Plots with black and red markers correspond to the throughputs obtained by 3 uploading and 3 downloading connections, respectively, with a wireless channel error probability of p w = 0.3. Throughput obtained by different STAs are marked with different symbols Observation of time scale difference in the evolution of congestion windows as in experimental (NORNET) with ns-2 simulations results. The figure is the congestion window evolution of Cellular and Wi-Fi subflow when the average packet loss probability is 20% with unbounded AP and BS buffer using our ns-2 simulation testbed xii

20 List of Tables 2.1 Summary of well-known TCP Models Model Validation Techniques Literature on performance modeling of TCP over infrastructure WLAN Parameter Setting B r and p b (r) for r = 1 with N u = N d = 5 and N u = N d = 1 for different channel error probabilities, p w B r and p b (r) for r = 2 with N u = N d = 5 for different channel error probabilities, p w Simulation Parameters System Parameters xiii

21 Notation N u Total number of uploading STAs N d Total number of downloading STAs K Maximum number of (re)transmissions per packet, K = 7 M Maximum number of failures up to which the backoff window is doubled, M = 5 p w Probability that a packet transmission fails due to wireless error. W max Maximum TCP receive window B ap AP buffer size (in packets) r Weight of fairness, desired ratio of aggregate download throughput to upload throughput π(d, u) Stationary probability that there are u active uploading and d active downloading STAs Ps u Probability that the success belongs to upload STAs given that it is a success slot Ps Ap,Data Probability that the success belongs to the AP given that it is a success slot with AP holding TCP DATA at the HOL position p lu Probability that a TCP DATA packet is discarded by an upload station p ld Probability that a TCP DATA packet is discarded by the AP p d Probability that a TCP DATA packet is discarded h Fraction of services by the AP that are TCP DATA packets; also equal to the probability that the HOL packet at the AP is a TCP DATA packet (used in chapter 3) h r Fraction of services by the AP that are regular TCP DATA packets; also equal to the probability that the HOL packet at the AP is a regular TCP DATA packet(used in chapter 4) E[C] Expected number of failures in a renewal cycle E[A] Expected number of attempts in a renewal cycle E[W d ] Expected congestion window for a download connection E[W u ] Expected congestion window for an upload connection Θ u Unrestricted aggregate TCP throughput of upload connections Θ d Unrestricted aggregate TCP throughput of download connections E (d,u) [X] Expected duration of a cycle γ u Probability of transmission failure for an upload station γap Data Probability of transmission failure for the AP p b Probability of buffer overflow at the AP or the probability that a packet arriving to the AP is blocked B r Desired AP buffer size (in packets) Θ u Aggregate TCP Throughput of upload connections (in packets/second) Θ d Aggregate TCP Throughput of download connections (in packets/second) M m Total number of multipath TCP users Total number of regular (i.e., single path) TCP users M r xiv

22 T d B bs B ap r W max X c X c π c (k) c (X c ) RT T w X w,k X w,k π w,k (l) w,k (X w,k ) λ w,k p m,k π a (u) E u [Y ] Θ w θ c p c p m p r θ m Round trip propagation delay (in seconds) over the cellular path Round trip propagation delay (in seconds) over the WiFi path Buffer size (in packets) at the cellular Base Station (BS) per user Buffer size (in packets) at the WiFi Access Point (AP) shared between all users Service rate (in packets/sec) of the cellular base station per user Receiver advertized maximum window (in packets) The congestion window pertaining to the cellular path of a multipath TCP connection The integer part of Xc ; to be called the cellular window Stationary probability that the cellular window X c is equal to k Increment of the cellular congestion window X c for every received acknowledgment (ACK) Expected value of the round trip time over the WiFi path The congestion window pertaining to the WiFi path of a multipath TCP connection, given that X c = k The integer part of Xw,k, given that X c = k; to be called the WiFi window Stationary probability that the WiFi window X w,k is equal to l, given that X c = k Increment of the WiFi congestion window X w,k for every received ACK, given that X c = k Conditional sending rate of a multipath TCP connection over the WiFi path, conditioned on X c = k Conditional buffer loss probability perceived by a multipath TCP connection over the WiFi path, conditioned on X c = k Stationary probability that there are u active WiFi users Expected value of the cycle duration with u active WiFi users, where Y is the random cycle duration representing the time between two consecutive successful transmissions on the shared WiFi channel Aggregate throughput of the WiFi system, which is the total rate (in packets/sec) at which the AP serves the packets of all the connections/users Throughput (in packets/sec) obtained by a multipath TCP user over the cellular path Average buffer loss probability perceived by a multipath TCP connection at the BS buffer Average buffer loss probability perceived by a multipath TCP connection at the AP buffer Average buffer loss probability perceived by a regular TCP connection at the AP buffer Throughput (in packets/sec) obtained by a multipath TCP user over the WiFi path xv

23 θ r Throughput (in packets/sec) obtained by a regular TCP user (over the WiFi path) n r (t) Number of regular TCP users at time t n m (t) Number of multipath TCP users at time t x r (t) Congestion Window of individual regular TCP user at time t x m (t) Congestion Window pertaining to the WiFi path of individual multipath TCP user at time t x c (t) Congestion Window pertaining to the cellular path of individual multipath TCP user at time t θ r (t) Throughput in packets/sec of individual regular TCP user at time t in the WiFi θ m (t) Throughput in packets/sec of individual multipath TCP user at time t in the WiFi θ c (t) Throughput in packets/sec of individual multipath TCP user at time t in the Cellular B ap AP buffer size (in packets) B bs BS buffer size (in packets) Q(t) Total number of packets in the AP buffer at time t. q r (t) Number of packets in the AP buffer at time t belonging to a regular TCP user. q m (t) Number of packets in the AP buffer at time t belonging to a multipath TCP user. q c (t) Number of packets in the BS buffer at time t belonging to a multipath TCP user. h r (t) Fraction of packets in the AP buffer at time t belonging to a regular TCP user. p r (t) Packet Loss probability experienced by the regular TCP packets at time t. p m (t) Packet Loss probability experienced by the multipath TCP packets on the WiFi path at time t. p c (t) Packet Loss probability experienced by the multipath TCP packets on the cellular path at time t. p w Probability that a transmission of a data packet fails due to wireless channel errors. RT T (t) Round trip time over the WiFi path at time t. RT T c (t) Round trip time over the cellular path at time t. xvi

24 Chapter 1 Introduction The Transmission Control Protocol (TCP) has been the core transport protocol in the Internet for more than two decades. The majority of data traffic in the Internet is still carried by TCP [1]. Despite various refinements over the years, its fundamentals have remained essentially unchanged. But in recent years real momentum has been building for a major rethink of TCP and wireless standards. This is reflected, for example, in the establishment of a group to extend TCP for multipath operation and the development of multipath TCP [2], being standardised by the Internet Engineering Task Force (IETF) and used by Apple, as well as in the recent establishment by European Telecommunications Standards Institute (ETSI) of a working group on next generation protocols for 5G (fifth generation) Cellular networks. Important trends include: (i) a move towards multipath operation, especially use of both WiFi and Cellular on mobile handsets, and (ii) concern to improve operation over heterogeneous wireless networks. Therefore, we are at the cusp of transformation of the TCP protocols and wireless communication systems. Internet users are increasingly becoming wireless and utilizing multiple network paths, often using reliable TCP connections over the Cellular and WiFi networks [3]. This creates demands as well as opportunities to investigate the performance, i.e., the Quality of Service (QoS) perceived by the Internet users while using last-mile wireless networks. Challenges of these systems involve the growing number of wireless devices and traffic [4] that require intelligent co-ordination, monitoring and management in a distributed manner, mainly, for reliable communications and fair sharing of the resources among users. In the view of above discussion, a number of challenges need to be resolved in order to realise all these changes. More specifically, we identify the following key challenges: 1

25 1. Variable path loss: The likelihood that a TCP packet will be lost while traversing a wired-cum-wireless network path is stochastic and time-varying as a result of buffer size, queueing, other flows sharing the path, wireless loss and insufficient retransmissions. When packets are sent via wireless paths they therefore can be easily lost and cannot always arrive at the destination. As a result the lost packets are needed to be retransmitted until they can be delivered to the destination. This demands subsequent TCP and link layer retransmissions of packets. The resulting impact of all types of losses and the variation in loss can be substantial, to the point where it largely undermines the throughput gain from the use of multiple paths. 2. Heterogeneity in path characteristics: When a TCP connection is first initiated the end hosts have little knowledge of the characteristics of the path. Typically, information is confined to that gathered from the initial connection handshake and perhaps also from some known facts, e.g. the Cellular path typically uses large buffers, resulting in long delays and low loss rates, whereas the WiFi path might have smaller delays and higher loss rate [5]. As the connection proceeds, feedback is obtained from packet transmissions but this feedback is delayed due to queuing and propagation in the network. Short connections (short-lived flows) thus have limited information as to path characteristics. However, the longer connections (long-lived flows) need to adapt with the heterogeneous path characteristics as data is being transmitted, which is the primary focus of this thesis. 3. Resource allocation: Resource sharing in today s Internet is primarily achieved through the use of a single path transport protocol, called the regular TCP [6]. Different loss- and delay-based versions of TCP provide effective congestion control and fair bandwidth sharing in the Internet in a distributed manner, ensuring its stability and efficiency. In the absence of any QoS guarantees namely, the best-effort Internet, the fairness of bandwidth sharing provided by TCP has played a crucial role in ensuring its widespread acceptance as the de facto transport protocol [7]. Therefore, understanding the issues of coexistence, such as intra- and inter-protocol fairness and friendliness towards the standard TCP, has ever been an integral challenge in the design of transport protocols [8 11]. Looking ahead, as multipath TCP flows become more prevalent, the potential exists for the network to assist these flows so as to improve overall network performance, dynamic resource allocation and user QoS. For example, when some 2

26 users have good WiFi or Cellular connections and others have poor connections it may be beneficial to reassign WiFi or Cellular capacity from the former to the latter in some way. Furthermore, it is well-known that the performance of regular TCP tends to be poor when used over lossy wireless networks [12, 13]. In particular, the erratic nature of the last-mile wireless channel can lead to several back-to-back (or correlated) losses of packets. To ensure reliability, the lost packets must be detected and retransmitted by the sender which involves long delays due to RTTs. Unlike multipath TCP, regular TCP cannot take advantage of the multiple network interfaces in modern user devices. Multipath TCP could transfer data to and from such devices through multiple interfaces and using multiple available paths in the Internet, thus providing improved performance and robustness against possible failure of any path. As mentioned earlier, the IETF has set up a working group that is currently actively progressing towards the enhancement of multipath TCP [14 16]. With multipath TCP, a user can dynamically add multiple network addresses without disruption and utilize multiple network interfaces and paths, if available. Experimental studies have indicated that, for long file transfers over paths of comparable quality, multipath TCP is capable of providing significant gains over regular TCP [17 20]. In this thesis we address the key challenges discussed above as follows. We develop analytic models, analyse the performance using mathematical theories, and design the control mechanisms to overcome the aforementioned challenges. Our focus is on modelling the impact of the last-mile wireless networks on the QoS experienced by the mobile Internet users using long-lived TCP connections. The purpose of performance modeling of the multipath TCP dynamics over wireless networks in this thesis is to develop simple and accurate analytical models that can explain the behaviour, as observed in the detailed computer-based simulations and also reported in the literature, using real testbed measurements for long-lived connections. While developing mathematical models, our focus is on the analysis for predicting the key performance metrics, such as throughput, loss and delay, so as to validate the accuracy of the developed models. We evaluate the developed models extensively to test the validity of the assumptions made during their design, gain confidence in their practical utility over live Internet paths, and to increase usability. Such analysis can provide additional insights into the stability and fundamental properties for the control of system dynamics. 3

27 The main step towards the modeling study in this thesis is to obtain an accurate model for the system dynamics depending on the wireless networks scenario for multiple long-lived connections. The model essentially captures the congestion control, wireless access technique, carrier sensing and channel errors perceived by the users while using the heterogeneous Cellular and WiFi networks simultaneously. We investigate the behaviour of networks of multipath connections and ways in which network operators can help manage resource allocation. One major focus is the provision of path analytics to assist long-lived TCP connections, e.g. by providing path-based loss management with admission control techniques. We propose network-assisted QOS by controlling the buffering policies at the last-mile WiFi/Cellular networks. A second major focus is dynamic resource allocation at the wireless edge to optimize the usage of the diverse characteristics of WiFi and Cellular with respect to QoS. Specifically, the fundamental heterogeneous characteristics of WiFi and Cellular last-miles considered in this thesis are: i) in Cellular network, each user device is allocated a separate channel, however, in WiFi network, the common medium has to be shared by multiple users, leading to contentions for medium access and collisions; and ii) in Cellular base station, it is often the case that each flow is allocated a separate buffer, however, the same bottleneck buffer at the WiFi AP has to be shared by multiple TCP users. We evaluate the proposed control algorithms extensively. We validate the numerical results with simulations and gain confidence in their practical usability over the live WiFi/Cellular networks. 1.1 Thesis Statement In view of above discussion, the fundamental research question of this thesis is the thesis statement, i.e. Q How do the heterogeneous last-mile wireless access technologies affect the performance of TCP? Moreover, Q is a broad research question, which consists of a set of several research questions based on factors including the impacts on different types of TCP protocols (regular vs multipath), heterogeneous wireless networks (WiFi vs Cellular), wireless channel conditions, network buffers, number of users and number/type of transmission paths. 4

28 TARGET Impact of Last-Mile Wirless on Regular and Multipath TCP SCENARIOS Regular TCP WiFi or Cellular (Traditional scenario) Chapter 3 Multipath TCP WiFi and Cellular (Future scenario) Chapter 4 Coexisting Regular and multipath TCP WiFi and Cellular (Current scenario) Chapter 5 CHALLENGES Buffer Loss Channel Loss Joint Impact Unfairness Path charcterstics Joint Evolution Time Scales Unfairness Variable users Transient dynamics Coexistence Stability OUTCOMES Stochastic Model (Stationary) Fluid Model (Transient) Fairness Solution (Static) Dynamic Control of Fairness Figure 1.1: Structure of the Thesis In this thesis, we focus on the three main directions of the aforementioned research question as illustrated in Fig Fig. 1.1 represents the overall structure and directions of the research consisting of scenarios with challenges, outcomes and the target of the different directions of this thesis. It can be observed from the Fig. 1.1 that the structure of this thesis consists of separate design spaces with three different network scenarios, identification of the research challenges and problems in each of the scenario, development of the corresponding models, and finally design of solutions to the identified problems by using the deliverable from the developed models. The three major research questions of the thesis are (see Fig. 1.1) Q1 What are the impacts of correlated buffer losses and shared channel characteristics on the performance of TCP over WiFi networks? Q2 How well multipath TCP performs with the heterogeneous combination of Cellular and WiFi paths? Q3 Do regular and multipath TCP co-exist well in such heterogeneous networks? 5

29 In the following sections, we will discuss why the three main directions of our research are important and how stochastic and analytical methods can be applied to investigate the research questions. Furthermore, we will list our novel contributions in each direction of the research. 1.2 Regular TCP Performance over WiFi With the growing presence of WiFi and Cellular last-mile Internet access, the evolution towards Internet of Things (IoT) has been already evident [21]. About 90% of the DATA traffic in the Internet today is carried by the TCP [1], and a majority of that traffic will be preferably transferred via a path with WiFi last-mile which is significantly faster and cheaper than a Cellular connection [22]. An in-depth understanding of TCP dynamics over WiFi networks is thus essential to effectively design, deploy and manage a large number of wireless devices, and analytical modeling can provide such important insights. WiFi networks, or more precisely, IEEE wireless local area networks (WLANs), often operate in a so-called infrastructure mode, as illustrated in Fig 1.2. In infrastructure WiFi [23], user devices or stations (STAs) associate with an Access Point (AP) and transceive the DATA using a variant of the Carrier Sense Multiple Access (CSMA) Medium Access Control (MAC) protocol. An accurate prediction of the performance of regular TCP over infrastructure WiFi Internet access will be highly useful for efficient selection of low-cost paths. For example, multihomed STAs that are in the vicinity of two or more APs may dynamically switch their associations between the APs based on some analytically predicted performance metrics, and thereby, achieve optimal performance. The performance of TCP over wireless networks has been widely studied in the literature using simulations, analytical methods and real-world experiments. Detailed surveys can be found in [24], [25], and [26]. One class of analytical models for such networks studies the detailed dynamics of TCP with an abstraction of the lower layers. This is done via approximating the actual packet loss process by an analytically tractable loss model. The analysis with an i.i.d. packet loss model can be found in [27] and that with stationary random losses is developed in [28]. Abouzeid et al. [29] analyze the performance of TCP Reno accounting for both wireless and the buffer losses. A comprehensive model for the performance of TCP NewReno with Cellular last-mile access can be found in [30], where the impact of auto- and cross-correlations between wireless and buffer losses is investigated. 6

30 To Internet AP STA Wireless link Wireline link Figure 1.2: An infrastructure WLAN consisting of an AP, uploading STAs and downloading STAs. The STAs communicate with servers in the Internet through the AP using TCP connections. Nevertheless, none of the aforementioned work takes into account the impact of the MAC access layer (e.g., WiFi channel contention and collisions) on the performance of TCP. Our goal is to understand the performance of TCP over an infrastructure WiFi, as illustrated in Fig 1.2 where the MAC layer plays an important role, and one cannot approximate the lower layers simply by a packet loss model. Modeling the joint impact of buffer and wireless losses over WiFi last-mile access is significantly more challenging than that over Cellular last-mile for the following two reasons. 1. Shared Channel vs Dedicated Channel: Each user device is allocated a separate channel in Cellular. However, in WiFi, the common medium has to be shared by multiple users, leading to contentions for medium access and collisions. Unlike in Cellular networks, where the impact of other users can be modelled through an interference power, the time sharing of the common medium in WiFi networks requires a detailed and careful modeling of the contention for medium access. 2. Shared Buffer vs Dedicated Buffer: It is often the case that each TCP user is allocated a separate buffer in Cellular. In WiFi, however, the same bottleneck buffer at the AP has to be shared by multiple TCP users. Unlike in Cellular 7

31 networks, where there is no impact of other users, the space sharing of the common AP buffer in the WiFi network requires a detailed model for tracking the fraction of packet services from the AP belonging to the user. variable conditions such as number of flows sharing the WiFi, retransmission limit and magnitude of wireless errors in the Cellular and WiFi channel, will affect the TCP level packet loss probability. Consequently, those losses are more likely to be correlated. Correlated losses are also introduced due to drop-tail buffer at the AP and BS, which is the most common buffering policy implemented in the AP and BS buffers. for example, there can be the case where some packets may fail due to buffer overflows and others which successfully enter in to the buffer, may fail due to wireless channel errors [31, 32] Contributions Our major contributions in the direction of regular TCP performance over WiFi are as follows. 1. We develop a model that can capture the joint impact of buffer and channel losses on the performance of TCP enriched with comprehensive treatment for the MAC discards, i.e. packet losses due to the combination of wireless errors and MAC collisions. The interplay of both uploading and downloading users in our model is non-trivial because of the coupling between the two directional TCP traffic influenced by the WiFi contention access mechanism. 2. We discover that wireless channel errors cause a throughput unfairness that is opposite by nature to the well-known unfairness due to buffer overflows. Our analytical model provides the means to gain detailed insights into this unfairness, as well as an explanation of this behavior. 1.3 Performance Analysis of Multipath TCP As discussed earlier, with the growing presence of smart-phones and the ensuing surge of mobile Internet traffic [3], last-mile Cellular networks are becoming overwhelmed. The situation is anticipated to worsen in the near future [4]. A recent approach adopted to alleviate this problem is to offload the data traffic through WiFi networks [22]. A better approach could be to use an advancement of the TCP, the multipath TCP [2], which can perform parallel data transmission over the multiple 8

32 Figure 1.3: Multipath TCP users are downloading data from a remote server in the Internet using their WiFi and Cellular interfaces. The WiFi Access Point (AP) is shared by all multipath TCP connections. The Cellular paths of the multipath connections have dedicated buffer and wireless channel at the Base Station (BS). available paths. Since most smart-phones are equipped with the Cellular and the WiFi interfaces, multipath TCP can significantly improve the reliability and quality of Internet access through the smart-phones by simultaneously utilizing the Cellular and WiFi paths and dynamically balancing the traffic as the number of competing flows changes. The performance of multipath TCP over the last-mile wireless networks has been experimentally studied [5, 15, 33, 34] and useful insights have been gained. However, detailed investigations are necessary to quantify the impact of heterogeneous wireless access technologies on each individual path and their coupling. Moreover, a thorough understanding of the impact of heterogeneous wireless access technologies is essential to further improve the multipath congestion control algorithms, and analytical modeling could provide such important insights. Most of the recent analytical work on multipath transport [35 37] is primarily concerned with applying fluid models to derive improved versions of congestion control algorithms. However, these works [35 37] do not analyze the packet-level queueing 9

33 and wireless medium access layer contentions that determine the Round Trip Times (RTTs) and packet loss probabilities. In particular, it has been reported that often the Cellular paths are of low bandwidth and high delay, whereas the WiFi paths are of high bandwidth and low delay [5, 15]. The recent work [5] provides only a partial analysis of the coupling between the Cellular and WiFi paths and does not provide a method for analytically predicting important performance metrics such as throughputs and buffer losses. Our goal is to provide complete analysis of the scenario as illustrated in Fig. 1.3 under the congestion control mechanism of multipath TCP Linked Increases Algorithm (lia), which is the standardized multipath TCP protocol by IETF Contributions Our major contributions in the direction of performance analysis of multipath TCP are as follows. 1. We develop an analytical model for long-lived multipath TCP connections over heterogeneous Cellular and WiFi networks as illustrated in Fig. 1.3 by using the time scale difference analysis. 2. We discover that if a mixture of regular and multipath TCP connections share a WiFi access point with a droptail buffer, then the regular TCP connections steal almost all the WiFi bandwidth away from the multipath TCP connections. We expand our analytical model to the case where a mixture of regular and multipath TCP connections. Our analytical model in the mixed case case, when solved, provides the appropriate value of the equal admission probability that must be used to alleviate the throughput unfairness. 1.4 Co-existence of Regular and Multipath TCP Fair sharing of network resources has been one of the main objectives in the early design and development of multipath transport protocols [38, 39], and ensuring the friendliness of multipath TCP towards the standard TCP has been an essential design goal in the development of multipath TCP [2, 36, 40]. While the coexistence among different versions of regular (or, single path) TCP users is rather well understood [41 46], the coexistence between regular and multipath TCP users is a relatively new topic [36,40,47 51]. The study in [36] argues that multipath TCP lia is not Pareto-optimal as it reduces the throughput of regular TCP users without increasing 10

34 the throughput of multipath TCP users. Peng et al. [40] show that there exists a design trade-off between the responsiveness and TCP friendliness, and propose an enhanced algorithm called Balia (Balanced lia) that balances the responsiveness and TCP friendliness. The friendliness in the reverse direction (i.e., of regular TCP towards multipath TCP), however, has not received sufficient attention in the literature so far. An exception is [20] wherein the authors point out that the conservative behavior of multipath TCP cannot compete with regular TCP when its subflows are not sharing the common bottleneck. Any such adverse impact of regular TCP on multipath TCP could have major repercussions on the wide acceptance of multipath TCP, or on its evolutionary stability [52]. When homogeneous single path TCP sources share a network, one can always associate an appropriate underlying utility function. However, with heterogeneous single path and/or multipath transport protocols sharing a network, in general, one may not interpret the sources to be maximizing any underlying utility functions [53, 54], [40]. In particular, with heterogeneous protocols, the bandwidth allocation depends on the queue management in place [54]. Therefore, one may not directly perform utility optimization to design multipath source control algorithms and achieve a desired notion of fairness when sharing the network with regular TCP sources Contributions In the direction of co-existence of regular and multipath TCP, we have the following contributions. 1. We develop a dynamic transient model and study the coexistence of regular and multipath TCP users sharing a WiFi access point. 2. Based on the closed form expression derived using the analytical model under equal loss condition, we propose a real-time algorithm that continuously monitors the deviation in buffer occupancy at the AP to ensure fairness. 1.5 Organization of the Thesis The thesis is organized as follows: 1. Chapter 2 is organized as follows. Sec. 2.1 provides preliminary background related to TCP protocols and WiFi standards. Sec. 2.5 reviews the existing 11

35 models of TCP over wireless, and the existing approaches of performance analysis. 2. Chapter 3 is organized as follows. In Sec. 3.2, we describe our network scenario in detail in order to motivate our modelling assumptions in Sec An analytical model using the fixed point approach is developed in Sec We find that the wireless channel errors and buffer overflows both lead to throughput unfairness, but that they do so in the opposite direction on the up and down links, respectively. In Sec. 3.4, the model is validated using simulations and the new type of unfairness is discussed in detail. Two methods to achieve throughput fairness are proposed and their effectiveness is studied in Sec The Chapter concludes in Sec This Chapter of the thesis is based on the studies in [55]. 3. In Chapter 4, we develop a comprehensive analytical model for multiple longlived multipath TCP connections downloading content from a remote server in the Internet using parallel paths with WiFi and Cellular last-miles. In Sec. 4.2, we describe our network scenario in detail that forms the basis for the simplifying assumptions that we make in developing our analytical model. An analytical model for a set of multipath connections sharing an AP with a droptail buffer is developed in Sec In Sec. 4.5, we report a new type of throughput unfairness that we observe when a mixture of regular and multipath TCP connections share an AP with a droptail buffer. To tackle this problem, we present two simple solutions utilizing our analytical model and achieve fairness. In Sec , we expand our analytical model for the multipath-tcp-only case to the mixed case with packet admission control at the AP buffer. In Sec , we apply the admission probability obtained with our analytical model in two different ways and significantly reduce the throughput unfairness. This Chapter of the thesis is based on the studies in [56]. 4. Chapter 5 is organized as follows. We explain the network scenario and assumptions in Sec We develop a combined packet and fluid level analytical model in Sec The strategy for fairness using loss equalization and the algorithm is detailed in Sec The numerical results of the algorithm are validated with NS-2 simulations in Sec We prove in Sec that the proposed algorithm is stable by using the Lyapunov s stability theorem of dynamic systems. The algorithm results are validated with NS-2 simulations. We conclude the Chapter in Sec This Chapter of the thesis is based on the studies in [57]. 12

36 5. In Chapter 6, we conclude the thesis with highlights of the major contributions, shortcomings and shed light towards novel future research directions. This Chapter briefly summarizes the thesis, provides overall discussion and open research issues in the considered domain. 6. The appendix includes the detailed proofs and mathematical analysis of the results presented in Chapters 2 and Publications Included in the Thesis J1 Pokhrel, Shiva Raj, Manoj Panda, Hai L. Vu, and Michel Mandjes. TCP Performance over Wi-Fi: Joint Impact of Buffer and Channel Losses. IEEE Transactions on Mobile Computing 15, no. 5 (2016): J2 Pokhrel, Shiva Raj, Manoj Panda and Hai L. Vu. Analytical Modeling of Multipath TCP over Last-Mile Wireless. IEEE/ACM Transactions on Networking, (2017): Accepted for Publication. 23 Jan 2017 J3 Pokhrel, Shiva Raj, Manoj Panda and Hai L. Vu. Fair Co-existence of regular and Multipath TCP over Wireless Last-Miles. Submitted to IEEE Transactions on Mobile Computing Not included in the Thesis J4 Panda, Manoj, Hai L. Vu, Michel Mandjes, and Shiva Raj Pokhrel. Performance analysis of TCP NewReno over a Cellular Last-Mile: Buffer and Channel Losses. IEEE Transactions on Mobile Computing 14, no. 8 (2015):

37 Chapter 2 Background and Literature Review This chapter provides the state-of-the-art for IEEE WiFi and Cellular network focusing on the TCP protocol. It includes the analytical models of the TCP protocols over wireless networks, and the existing approaches for the performance analysis and evaluations of WiFi and Cellular networks. It is well-known that the widely used technologies for the shared wireless channel access mechanism are the IEEE WiFi and Cellular networks [58]. Similarly, the dominant transport layer protocol for reliable data transfer is the TCP protocol. Therefore, this work focuses on the TCP performance over WiFi and Cellular networks. We first summarise the standards and the working mechanisms of TCP and WiFi/Cellular networks in Sec. 2.1 and 2.2. The analytic modelling methods for quantifying the TCP performance and dynamics are discussed in Sec. 2.4 and 2.5. These techniques develop the performance metrics in terms of the relevant parameters like congestion window, packet loss and queuing delay in the networks (Sec. 2.5). Note that the existing models provide accurate predictions in case of wired networks, where these parameters are almost always constant and known quantities. However, most of these models are not applicable in the case of wireless networks because the parameters such as packet loss and delay in wireless networks are highly fluctuating and unpredictable, particularly, in Cellular and WiFi networks. Furthermore, the performance evaluation of multipath congestion control in lastmile wireless networks is more complex and challenging than that of regular TCP because the coupling of flow rates (using windows) across the available paths are being affected by the heterogeneous and fluctuating characteristics (loss and delay) of the wireless paths [5,15,34]. To help the reader grasp the directions on the modeling of multipath TCP protocols over the last-mile wireless networks and the coexistence of regular with multipath, we will summarise the existing literature and models of multipath TCP. 14

38 2.1 Last-Mile Wireless Access Networks This section provides a background of both Cellular and WiFi access technologies relevant to our thesis. Further information can be found in [23, 59 62] Cellular Access Technologies The Global System for Mobile Communication (GSM), the Cellular standard was originally described as a circuit switched network optimized for full duplex voice telephony. It was the most popular wireless access technology first defined in 1990 by ETSI for the second generation (2G) digital Cellular networks [59] to replace first generation (1G) analogue systems such as NMT (Nordic Mobile Telephony) and TACS (Total Access Communication System). IS-95 (Interim Standard 95) is another 2G mobile wireless technology standard that uses CDMA (Code Division Multiple Access) [63], a multiple access scheme for digital radio to send voice, data and signalling data between mobile users and cell transmitting station. The GSM standard was enhanced to include packet data transport via GPRS (General Packet Radio Service). Packet data transmission speeds were later increased via EDGE (Enhanced Data rates for GSM Evolution). In 2003, EDGE or EGPRS (Enhanced GPRS) [59] was introduced which is a digital wireless technology that allows improved data transmission rates. Interestingly, no hardware/software upgradation were needed in the core network, however, EDGE transceiver units were required to be installed. EDGE allows data rates of 59.2 kb/s per time slot [59]. The GSM standard was succeeded by the third generation Universal Mobile Telecommunications System (UMTS) standard developed by the Third Generation Partnership Project (3GPP). GSM networks evolved further as they begin to incorporate fourth generation (4G) LTE-Advanced standards [62]. The pictorial representation of the evolution of Cellular mobile standard from 2G to 5G is illustrated in Fig. 2.1 [64]. UMTS [65] specifies an entire network system, covering the radio access network UTRAN (UMTS Terrestrial Radio Access Network), the core network (Mobile Application Part) and the authentication of mobile users via SIM cards. A number of channel types exist divided into physical channels, transport channels (further categorized into common transport channels and dedicated transport channels) and logical channels. Data packets can be sent using a contention based uplink channel (Random Access Channel, RACH) or a common downlink channel (Forward Access Channel, FACH) using a common spreading code [65]. 15

39 Figure 2.1: Evolution of cellular mobile standard from 2G to 5G. CDMAone and CDMA2000 [66] form a parallel development track to GSM and UMTS using Code Division Multiple Access as channel access method and a duplex pair of 1.25 MHz radio channels. CDMA2000 being a successor of CDMAone was standardized by 3GPP2 and upgraded from the first version to the Evolution-Data Optimized (EV-DO) versions offer data rates of 3.1 Mbps and 1.8 Mbps in the downlink and uplink directions. With a hardware upgrade some versions can even support sppeds of 14.7 Mbps and 5.4 Mbps in the downlink and uplink directions. LTE is a suitable standard for last-mile data access in the Cellular network technology with a set of enhancements to the UMTS which was introduced in 3GPP Release 8 [67, 68]. LTE substitutes the WCDMA transmission scheme of UMTS so that Orthogonal Frequency-Division Multiple Access (OFDMA) is used for downlink while Single-carrier FDMA (SC-FDMA) is used for uplink transmission. OFDM is a type of modulation scheme that is used as a digital multi-carrier modulation method where a number of closely spaced orthogonal sub-carriers are used to carry data. The data is divided into several parallel data streams or channels, one for each sub-carrier [62]. A dynamic resource allocation in LTE is achieved through real-time assignment of sub-carriers to a specific user. Fourth generation (4G) [69] standard as defined by the International Telecommunication Union (ITU) states that LTE has the potential to be a true 4G Cellular technology. LTE Release 10 is LTEAdvanced [70], which is not a new radio-access 16

40 technology, but simply a name given to LTE Release-10 and beyond [70]. ETSI has also established a working group on next generation protocols for 5G Cellular networks [71] Cellular Architecture In this Section, we discuss only the essential units of Cellular networks with particular importance to our thesis as shown in Fig. 2.2(see [59] for further detailed information). The GSM network can be broadly divided into the following parts. The Mobile Station (MS) represents the mobile user with subscriber identity module (SIM). The Base Station Subsystem (BSS) consists of the Base Switching Centre (BSC) and Base Transceiving Station (BTS). The BSC controls a group of BTSs, and manages the radio resources and controls items such as handover within the group of BTSs. It communicates with the BTSs over what is termed the Abis interface. The BTS used in a GSM network comprises the radio transmitter receivers, and their associated antennas that transmit and receive to directly communicate with the mobiles. The Network Switching Subsystem (NSS) consists of the Main Switching Centre (MSC) and its functional units as shown in Fig The MSC acts like a normal switching node within a PSTN or ISDN, but also provides additional functionality to enable the requirements of a mobile user to be supported. These include registration, authentication, call location, inter-msc handovers and call routing to a mobile subscriber. The Operation Support Subsystem (OSS) is for operation and maintenance of the system. The OSS is the element within the overall GSM network architecture that is connected to components of the NSS and the BSC. It is used to control and monitor the overall GSM network and it is also used to control the traffic load of the BSS. The additional components of the GSM architecture comprise databases and messaging systems functions (see Fig. 2.2). Namely, Home Location Register (HLR) database contains all the administrative information about each subscriber along with their last known location. 17

41 Mobile Station HLR EIR PSTN ISDN BTS BSC MSC OMC Mobile Station VLR AUC AIR INTERFACE (UM) ABIS INTERFACE A INTERFACE Figure 2.2: GSM architecture Visitor Location Register (VLR) contains selected information from the HLR that enables the selected services for the individual subscriber to be provided. Equipment Identity Register (EIR) is the entity that decides whether a given mobile equipment may be allowed onto the network. Authentication Center (AUC)is a protected database that contains the secret key also contained in the user s SIM card. SMS Serving Center (SMS SC) handles messages directed in different directions. Gateway MSC (GMSC)is the in-charge of obtaining the MSRN (Mobile Station Roaming Number) from the HLR based on the MSISDN (Mobile Station ISDN number, the directory number of a MS) and routing the call to the correct visited MSC. The MS and the BSS communicate across the Air interface. It is also known as the Um interface or the radio link. The BSS communicates with the Network Service Switching (NSS) center across the A interface. In a GSM network, a cell is the basic service area; normally, one BTS covers one cell. Each cell is given a Cell Global Identity, a number that uniquely identifies the cell. The MSC of GSM is connected with other communication networks such as PSTN and ISDN, and other service operators WiFi Architecture In the IEEE WLAN architecture, two major components are the Access Point (AP) and the stations (STAs) [23, 72]. STAs are usually the devices with wireless 18

42 Mobile Station Mobile Station Computer Mobile Station Mobile Station Computer Access Point Mobile Station Computer Computer Mobile Station Computer Computer ADHOC MODE INFRASTRUCTURE MODE Figure 2.3: Ad-hoc Mode and Infrastructure Mode of WiFi Architecture interface cards such as smartphones, tablets, laptops and computers. The AP is an entity that has all the functions of STA and also provides access to the distribution services, via the wireless medium for all of the associated STAs. There are two different modes of operation as shown in Fig Ad-hoc Mode: All STAs connect and communicate with each other directly. 2. Infrastructure Mode: All STAs connect and communicate with each other through the AP, which is a focus of this thesis WiFi Standards The IEEE WiFi standards define the specifications of the MAC and physical layer of WLANs. The IEEE standard was proposed in 1996 [73], updated in 1999 [72] and amended in 2007 [23] and 2012 [74]. There are many other amendments to this standard with an aim to improve the performance of WiFi by increasing capacity at the physical layer and quality of service at the MAC sublayer Physical Layer The physical (PHY) layer, the first layer of the OSI Model, deals with bit-level transmission between different devices and supports electrical or mechanical interfaces connecting to the physical medium for synchronized communication. It consists of a Physical Medium Dependent (PMD) sublayer and a Physical Layer Convergence Protocol (PLCP) sublayer. 19