Wireless Networks: Model and Optimization based approaches to Clock Synchronization, Random Access MAC and Video Streaming

摘要

We, via a model and optimization-based approach, address three issues related to wireless networks: clock synchronization, medium access control (MAC) and scalable video streaming.In Chapter 2 we develop, study and simulate a new model-based distributed network clock synchronization protocol. In a network of clocks, a given node is taken as reference and is associated with the time evolution t. We introduce and analyze a stochastic model for clocks, in which the relative speedup of a clock with respect to the reference node, called the skew, is characterized by an exponential transformation of an Orstein-Uhlenbeck process. We study the properties of our model, namely moment and sample path properties of the stochastic processes, and calculate its Allan variance. We show how our model can be used to translate the time of a clock to another clock's units. We study the problem of synchronizing clocks in a network, which amounts to estimating the instantaneousrelative skews and relative offsets, i.e., the differences in the clock readouts, by exchange of time-stamped packets between pairs of nodes in the network. Based on a stochastic model for delays, we derive a scheme for obtaining relative skew measurements in a communication link by sending two time-stamped packets from node i to node j in order to obtain a noisy measurement of their relative skew. We develop an algorithm for filtering relative skew measurements across a link (i,j) in order to estimate the logarithm of the relative skew. We study the properties of the algorithms and provide theoretical guarantees on their performance. We also develop an online, centralized, model-based, asynchronous skew estimation algorithm for optimal filtering of the time-stamps in the entire network, as well as an efficient distributed suboptimal scheme which demonstrates near-optimal performance in simulations. Furthermore, we study some implementation issues, and present a scheme for pairwise relative offset estimation given skew estimates. We use the distributed asynchronous algorithm to obtain nodal offset estimates from relative offset estimates. We combine our findings into developing a new protocol for clock synchronization, namely the Model-Based Clock Synchronization Protocol (MBCSP). We present a comparative simulation study of its performance versus the leading scheme by Solis et al. (2006); the results show that MBCSP performs better in terms of skew, offset and delay estimation. Finally, we have performed trace-driven simulation based on time-stamps obtained fromBerkeley motes. Our scheme outperforms that of Solis et al. by 45%, where we used the accuracy in predicting the receipt time-stamp at the sender as the clock synchronization metric.In Chapter 3, we study random access based MAC in the framework of network utility maximization (NUM). There has been much recent interest in protocol design for wireless networks based on maximizing a network utility function. A significant advance is theobservation that a decomposition of the Lagrangian suggests an approach where transmissions are scheduled to minimize back-pressure. However, a satisfactory MAC protocol that can realize such a scheduling algorithm is notably missing, and we develop one potential scheme. We present a candidate random access MAC protocol that extends an existing algorithm by Gupta and Stolyar (2006) in calculating the access probabilities. We also consider the online adaptation of access probabilities using local information about queue lengths and active links. We provide OPNET simulation results tocompare the performance of our scheme with the leading schemes. We estimate the capacity region of our scheme by simulation for various topologies and multiple flows. Our simulation studies indicate that our extension in conjunction with an implementation of back-pressure significantly outperforms the slotted-time algorithm of Gupta and Stolyar (2006).In Chapter 4, we present performance bounds for random access based MAC using carrier-sense multiple access (CSMA). In recent work, it was shown that a distributed CSMA-based MAC protocol is throughput-optimal which, in turn, implies that the class of controlled distributed random access MAC protocols can support the entire capacity region. It is challenging to study the performance of such schemes in terms of mean delays and compare it with some known results on the performance of centralized scheduling. We modify the model of Jiang and Walrand (2008) to obtain Markov chain models that incorporate the queue lengths as well as the information about the independent set, for single-hop networks. We show that the delay of the new models yields an upper bound on the delay of the original models. We derive upper and lower bounds on the mean total delay at the steady-state, and show that these bounds coincide with those for max-weight scheduling.Finally, we develop a method of deriving upper and lower bounds for random-access schemes by using linear programs (LPs). We present an optimization program for minimizing the upper bounds.In Chapter 5, we consider multihomed scalable video streaming systems where each video is concurrently transmitted over several access networks to a client. The problem is todetermine which video packets of a video stream to transmit, and associate each video packet with an access network, so that thevideo quality at the client is maximized under measured network conditions. We present a network model and a video distortion model to capture the network conditions and video distortion characteristics, respectively. We develop a mathematical formulation to find the streaming strategy for maximizing the average video quality at the client. While the formulation can be optimally solved using exhaustive search or dynamic programming,doing so takes a prohibitively long time, and is not practical for real-time video streaming servers. In order to efficiently solve the problem in real time, we propose several suboptimal convex problems along with two heuristic algorithms. We conduct extensive trace-driven simulations to evaluate the algorithms using real network conditions and actual scalable video streams. We compare our algorithms against the rate control algorithmsdefined in the Datagram Congestion Control Protocol (DCCP) standard. The simulation results show that our algorithms significantly outperform current systems while beingTCP-friendly. For example, compared to DCCP, our algorithms achieve at least 10 dB qualityimprovement and result in up to 83% packet delivery delay reduction. Finally, we study the trade-off between efficiency and optimality: One of the heuristic algorithms runs faster and is suitable for large-scale streaming systems, while the other one achieves better video quality and is more appropriate for smallerstreaming servers. The convex programming approach demonstrates a good trade-off between running time and performance.