xml version 1.0 encoding UTF8 standalone no
record xmlns http:www.loc.govMARC21slim xmlns:xsi http:www.w3.org2001XMLSchemainstance xsi:schemaLocation http:www.loc.govstandardsmarcxmlschemaMARC21slim.xsd
leader nam 2200385Ka 4500
controlfield tag 001 002029587
005 20090918080630.0
007 cr mnuuuuuu
008 090918s2009 flu s 000 0 eng d
datafield ind1 8 ind2 024
subfield code a E14SFE0002938
035
(OCoLC)436998301
040
FHM
c FHM
049
FHMM
090
TK145 (Online)
1 100
Xu, Yangyang.
0 245
Comparison of routing and network coding in group communications
h [electronic resource] /
by Yangyang Xu.
260
[Tampa, Fla] :
b University of South Florida,
2009.
500
Title from PDF of title page.
Document formatted into pages; contains 60 pages.
502
Thesis (M.S.E.E.)University of South Florida, 2009.
504
Includes bibliographical references.
516
Text (Electronic thesis) in PDF format.
520
ABSTRACT: In traditional communication networks, information is delivered as a sequence of packets from source to destination by routing through intermediate nodes which only store and forward those packets. Recent research shows that routing alone is not sufficient to achieve the maximum information transmission rate across a communication network [1]. Network coding is a currently researched topic in information theory that allows the nodes to generate output data by encoding their received data. Thus, nodes may mix the input packets together and send them out as fewer packets. Potential throughput benefit is the initial motivation of the research in network coding. Group communications refers to manytomany communication sessions where multiple sources multicast independent data to the same group of receivers. Researchers always treat group communications as a simple problem by adding a super source which is connected to all the sources with unbounded capacity links.However, it cannot control the fairness between different sources in this method. Additionally, the method may be incorrect in some scenarios. In this research, we will present an example to illustrate that and analyze the reason for that. The maximum multicast throughput problem using routing only is NPcomplete. Wu et al. introduced a greedy treepacking algorithm based on Prim's algorithm as an alternate suboptimal solution [2] This algorithm is modified in this work for group communications problem with routing in undirected networks. The throughput benefit for network coding has been shown in directed networks. However, in undirected networks, researchers have only investigated the multiple unicast sessions problem and one multicast session problem. In most cases, network coding does not seem to yield any throughput benefit [3] [4]. Li et al.introduced a cflow algorithm using linear programming to find the maximum throughput for one multicast session using network coding [3] We adapted this algorithm for group communications with network coding in undirected networks to overcome the disadvantage of the traditional method. Both algorithms were simulated using MATLAB and their results were compared. Further, it is demonstrated that network coding does not have constant throughput benefit in undirected networks.
538
Mode of access: World Wide Web.
System requirements: World Wide Web browser and PDF reader.
590
Advisor: Ravi Sankar, Ph.D.
653
Optimization
Throughput
Undirected networks
Subgradient algorithm
Multicast protocols
690
Dissertations, Academic
z USF
x Electrical Engineering
Masters.
773
t USF Electronic Theses and Dissertations.
4 856
u http://digital.lib.usf.edu/?e14.2938
PAGE 1
Comparison of Routing and Network Coding in Group Communications b y Yangyang Xu A thesis submitted in partial fulfillment of the requirements for the degree of Master of Science in Electrical Engineering Department of Electrical Engineering College of Engineering University of South Florida Major Professor: Ravi Sankar Ph.D. Jing Wang, Ph.D. Xiang dong Hou Ph.D. Date of Approval : March 2 4 200 9 Keywords: Optimization, Throughput Undirected Network s Subgradient Algorithm, Multicast Protocols Copyright 200 9 Y angyang X u
PAGE 2
ACKNOWLEDGEMENTS I would like to gratefully acknowledge the guid ance and support from my advisor Dr. Ravi Sankar for my study at USF. I would also like to thank Dr. Jing Wang and Dr. Xiang dong Hou for serving in my committee. I acknowledge the funding support from Raytheon Company and Florida High Tech Council for my research. I also want to thank Scalable Network Tech nologies for providing the QualN et simulation tool. I appreciate all the help coming from the iCONS group memb ers for my study and life in US. Finally, I thank my wife Jing, for her encouragement, support and lo ve
PAGE 3
i TABLE OF CONTENTS LIST OF FIGURES iii ABSTRACT v CHAPTER 1 INTRODUCT I ON 1 1.1 Background 1 1.2 Motivation and Goals 3 1.3 Outline of the Thesis 3 CHAPTER 2 UNICAST CAPACITY 5 2 2.1 Network Introduction 5 2.2 Cuts 7 2.3 Generic Augmenting Path Algorithm 8 CHAPTER 3 MULTICAST CAPACITY 1 6 3.1 Group Co mmunications 1 7 3. 2 Prim s Algorithm 1 7 3.3 Algorithm for Group Communications 20 CHAPTER 4 MULTICAST ROUTING PROTOCOLS 2 4 4.1 Comparison between Multicast and Unicast 2 4 4.1.1 Multicast s Advantages 25 4.1.2 Multicast s Disadvantage 26 4.2 Multic ast Models 26 4.2.1 ASM Model 2 7 4.2.2 SFM Model 2 7 4.2.3 SSM Model 2 7 4.3 IGMP 2 7 4.3.1 IGMPv1 2 8 4.3.2 IGMPv2 2 9 4.3.2 .1 Que rier Election Mechanism 30 4.3.2.2 Leav e Group Me chanism 30 4.3.3 IGMPv3 3 2 4.4 PIM Overview 33 4.4.1 PIM DM 3 4 4.4.2 PIM SM 3 5
PAGE 4
ii 4.5 Simulation of C urrent IP Multicast Protocols 3 6 CHAPTER 5 NETWORK CODING 3 9 5.1 Linear Network Coding 40 5.2 Practical Random Network Coding 4 2 5. 3 Subgradient Algorithm 4 4 5. 4 Algorithm for Network Coding in Group C o mmu nications 4 5 CHAPTER 6 SIMULATION AND COMPARISON OF RESULTS 5 1 CHAPTER 7 CONCLUSION AND FUTURE WORK 5 7 7.1 Summary 5 7 7.2 Recommendation s for Future Research 5 7 REFERENCES 5 9
PAGE 5
iii LIST OF FIGURES Figure 2.1 A Simple Network Topology 6 Figure 2 2 Min Cut (Node 1 is the Source and Node 4 is the Destination ) 8 Figure 2 3 Illustratin g the Generic Augmenting Path Algorithm 1 0 Figure 2 4 The Effect of Augmentation on Flow Decomposition 1 2 Figure 2.5 Flow Decomposition of the Solution in ( a ) Figure 2.4( a ) and ( b ) Figure 2.4( d ) 13 Figure 2.6 The Effect of Augmentation on F l ow Decomposition 13 Figure 3. 1 1 8 Figure 3. 2 1 9 Figure 3. 3 Example of our Algorithm for Group Communications 2 2 Figure 4.1 Example of Unicast 2 5 Figure 4.2 Example of Multicast 2 5 Figure 4.3 Example of IGMP v1 2 8 Figure 4.4 Example of IGMPv3 3 2 Figure 4.5 Network Topology for the Simulation of C urrent IP Multicast Protocols 3 7 Figure 5.1 Multicast with Routing 3 9 Figure 5.2 Multicast with Network Coding 40 Figure 5.3 Network Coding 4 1 Figure 5.4 Group Communications with Network Cod ing 4 6
PAGE 6
iv Figure 5. 5 Six Nodes Bi directional Network 4 7 Figure 6.1 Simulation N etwork Topology 5 2 Figure 6.2 Simulation Results of the Comp arison of Network Coding and Routing 5 3 Figure 6.3 Maximum Throughputs from Source Node 5 and 15 5 4 Figure 6.4 Maximum Throughputs from Source Node 1 2, 3, 4 and 5 5 5 Figure 6. 5 Maximum Throughputs from Source Node 1 to Node 10 5 5
PAGE 7
v Comparison of Routing and Network Coding in Group Communications Yangyang Xu ABSTRACT In traditional communication network s information is delivered as a sequence of packets from source to destination by r outing through intermediate nodes which only store and for ward those packets. Recent research show s that routing alone is not sufficient to achieve the maximum information transmissio n rate across a communication network [1] Network coding is a currently researched topic in information theory that allows the nodes to generate output data by encoding their received data. Thus, nodes may mix the input packets together an d send them out as fewer packets Potential throughput benefit is the initial motivation of the research in network coding. Group communication s refers to many to many communication sessions where multiple sources multicast independent data to the same group of receivers. Researchers al ways treat group communication s as a simple problem by adding a super source which is connected to all the sources with unbounded capacity links. However, it cannot control the fairness between different sources in this method. Additionally, the method may be incorrect in some scenarios. In this research, we will present an example to illustrate that and analyze the reason for that.
PAGE 8
vi The maximum multicast throughput problem using routing only is NP c omplete. Wu et al. introduced a greedy tree packing alg orithm based on Prim s algorithm as an alternate sub optimal solution [2] This algorithm is modified in this work f or group communications problem with routing in undirected network s The throughput benefit for network coding has been shown in directed networks. However, in undirected networks, researchers have only investigated the multiple unicast sessions problem and one multicast session problem. In most cases, network coding does not seem to yie ld any throughput benefit [3] [4]. Li et al. introduced a c flow algorithm using linear programming to find the maximum throughput for one multicast session using network coding [3] We adapted this algorithm for group communications with network coding in undirected network s to overcome the disadvantage of the traditional method. Both algorithms were simulated using MATLAB and their results were compared. Further, it is demonstrated that network codi ng does not have constant throughput benefit in undirected networks.
PAGE 9
1 CHAPTER 1 INTRODUCTION 1.1 Background Communication network is a directed graph model for the distribution of data from source to destination via intermediate nodes Routing i s the process of selecting paths in a network along which to send network traffic (voice/video/data packets) The network session is considered to be unicast when sending data from one source to one destination while it is termed multicast when sending dat a from one source to multiple destinations. The desired goal in any network is to fully utilize its capacity. Communication networks today share the same fundamental principle of operation. Independent data streams may share network resources, but the info rmation itself is separate. Routing, data storage, error control, and generally all network functions are based on this assumption. In the traditional communication networks, the intermediate nodes only copy and forward packets using the routing mechanism. Optimization problem for routing has been widely researched with graph theory. G eneric augmenting path al gorit hm is one of the basic algorithms used in unicast communications [ 5 ] Several other algorithms were developed to minim ize the computation complex ity For multicast communications Steiner tree algorithm is the state of the art for achieving optimal throughput. However, it is NP h ard.
PAGE 10
2 Several sub optimal algorithms have been developed to get the maximum throughput and the best ratio to the optimal t hroughput is 1/ 1.55 [ 6 ]. Ahlswede et al. showed that achieving multicast cap acity of a network requires network coding which allows intermediate nodes to generate output data by combining the received data [1] The output flow at a given node is obtained as linear combination of its input flows when we use linear network coding Linear network coding is sufficient to guarantee the multicast throughput to be the same as the maximum throughput from th e source to each individual destination in unicast session The initial expected benefit from network coding is the throughput benefit which has been shown in directed networks. However, from various studies reported in the past eight y ears, there is no throughput benefit in most of the scenarios in undirected networks. Li et al. showed that the network coding does not have any throughput advantage for one multicast session in 1,000 randomly generated undirected networks (the numbe r of links are less t han 35) [3] coding is not higher optimal throughput, but to facilitate significantly more efficient computation and implementation of strategies to achie Dougherty et al. concluded that there was no multiple unicast undirected network for which the coding capacity was large r than the routing capacity [ 4] However, group communications problem in undirected network s has not been studied
PAGE 11
3 1.2 Motivation and Goals The traditional method of solving the group communications has disadvantages. Also, there is nothing in the current literature on comparison of network coding and routing for group communications in undirected network s So, the primary goals of this thesis research are as follows: P resent the disadvan ta ges of traditional method for group communications and anal yze the reason for it. D evelop algorithms to get the maximum throughput for group communications i n undirected network s C ompare network coding and routing for group communications in undirected network s 1.3 Outline of the Thesis In this thesis, th e basic idea of unicast is first introduced in Chap ter 2. The max flow min cut theorem which is one of the most important theorems for network optimization problem is then presented Max flow min cut theorem is also the foundation of the new subgradient algorithm for network coding in group communications wh ich will be discussed later The generic augmenting path algorithm which is the basic algorithm for calculat ing the maximum flow in one unicast session with routing only (intermediate nodes only copy and forward packets) is presented as well Then, the multicast capacity problem is introduced in Chap ter 3. Steiner tree is the optim al solution for one multicast session problem. However, it is NP hard. So, Prim s algorithm and greedy algorithm designed by Wu for one multicast session problem is
PAGE 12
4 discussed [2]. Then, the proposed greedy algorithm for group c ommunications is that all sources can have the same fairness which other algorithms cannot guarantee In Chap ter 4, the real world Internet Protocol ( IP ) multicast with its advantage s and disadvantages comparing to IP unicast is introduced T he current widely used IP multicast protocols IGMP v1/v2/v3 for hosts to join and leave the multicast group and PIM SM/DM for routing scheme are discussed The simulation results for those protoco ls are compared to routing and network coding algorithms that are proposed in this thesis. In Chap ter 5, the network coding theory is introduced Then, the linear network coding and practical random network coding which can be the coding scheme for the s ubgradient algorithm are discussed The subgradient algorithm for one multicast session in undirected network designed by Li et al. [3] is then introduced. The proposed algorithm for group communications in undirected network s using network coding is prese all sources can have the same fairness. The simulation results of our greedy algorithm with routing and our subgradient algorithm with network coding for group communicat ions in undirected network s are presented in Chapter 6. The result s show that in most of the time, network coding does not have obvious throughput benefit for group communications in undirected network s
PAGE 13
5 CHAPTER 2 UNICAST CAPACITY 2.1 Network Int roduction Network is a directed graph model for the distribution of goods, data, or merchandise, etc., from their production centers, to their destination. Each directed edge has a limited capacity which is the maximum number of data that can be transmit ted through that channel per time period. The diagram indicates the capacity as a positive number associated with each edge. The actual number of data called the flow on that edge. It is a non negativ e number less tha n or equal to the capacity. Data cannot accumulate at a ny node; therefore the total in flow at each node must equal to the out flow at that node. The problem is to find the distribution of data that maximizes the net flow from s to t This can be modeled mathematically as follows. When the ed ges of a graph have a direction, the graph is called a directed graph or digraph. A network N is a directed graph with two special nodes s and t ; s is called the source and t is called the destination All other nodes in set I are called intermediate verti ces. The edges of a directed graph are ordered pairs (u,v) of vertices, which we denote by Each edge has a positive capacity which we denoted by If f is a real valued function defined on the edge set E and if we denote When we define If
PAGE 14
6 the function f() satisfies the following conditions, we define it as the flow of the network. Capacity constraint, ; Conservation condition, In general, there may be in edges as well as out edges at s The net flow from s to t will then be the out flow at the source minus the in flow. This is called the value of the flow, Any flow f that has maximum value for the network N is called a max flow of N This problem was first formulated and solved by Ford and Fulkerson [6], [17] Figure 2.1 A Simple Network Topology Since if this sum equals VAL(f). On the other hand, is the total out flow at Consider an out edge at v. Its flow contributes to It also contributes to If then will appear twi ce in the sum once for and once for and will therefore cancel. If then will appear in the summation as part of but will not be canceled by A similar argument holds if and Ther efore
PAGE 15
7 Eq. 2 .1 This says that the value of the flow can be measured across any edge cut such that and [ 7 ] In Chap ter 2 the maximum flow problem is considered subject to the following assumptions. Assumption 2.1 : The network is directed. Assumption 2 .2 : The network does not contain a directed path from node s to n ode t composed only of infinite capacity edges. Assumption 2 .3 : The network is delay less and error free. 2.2 Cuts Let N be a network with a single source s and a single destination t A cut in N is a set of edges of the form An s t cut is an cut with and The capacity of a cut is the sum of the capacities of its edges. We denote the capacity of cut K by cap K Thus, The s t cut whose capacity is the min imum among all s t cuts i s a minimum cut. Theorem 2.1 : Let be a s t cut and flow f Then If then f is a max flow and K is a min cut. (Max Flow Min Cut Theorem) As shown in Figure 2.2 the cut i s the min cut with node 1 as the source and node 4 as the destination. As a result, the maximum throughput from node 1 to node 4 will be 4+4 = 8 units/sec.
PAGE 16
8 Figure 2.2 Min C ut ( N ode 1 is the Source and Node 4 is the D estination) 2.3 Generic Augmenting P ath Algorithm Let us discuss one of the most intuitive algorithm s for solving maximum flow problem, the a ugmenting path algorithm. A directed path from the source to the sink in the residual network is referred as an augmenting path. We define the resi dual capacity of an augmenting path as the minimum residual capacity of any edge in the path. By definition, the capacity of an augmenting path is always positive. Consequently, whenever the network contains an augmenting path, we can send additional flo w from the source to the destination The generic augmenting path algorithm is essentially based on this simple observation. The algorithm proceeds by identifying augmenting paths and augmenting flows on these paths until the network contains no such path. The following is the description of the generic augmenting path algorithm.
PAGE 17
9 B egin f := 0; W hile G(f) contains a directed path from node s to node t do Begin I dentify an augmenting path P from node s to node t ; A ugment units of flow along P and update G(f) ; End E nd Algorithm 2.1 The maximum flow problem given in Figure 2.3 (a) illustrate s the algorithm. Suppose that the algorithm selects the path s 2 3 t for augmentat ion. The residual capacity of this path is This augmentation reduces the residual capacity of edge ( 2 3 ) to zero (thus we delete it from the residual network) and increases the residual capacity of edge (3, 2 ) to 4 (so we add this edge to the residual network). The augmentation also decreases the residual capacity of edge ( s 2 ) from 8 to 4, edge (3, t ) from 10 to 6 and increases the residual capacity of edge ( 2 s ) from 0 to 4 edge ( t 3) from 0 to 4 Figure 2.3 (b) shows the residual network after the first augmentation In the second iteration, suppose that the algorithm selects the path s 6 5 t The residual capacity of this path is Augmenting 4 unit s of flow along this path makes the residual network shown in Figure 2.3 (c). In the third iteration, the algorithm augments 4 unit s of flow along the path 1 6 5 t No w the residual network contains no augmenting path, so the algorithm terminates.
PAGE 18
10 (a) (b) (c) Figure 2.3 Illustrating the Gen eric Augmenting Path Algorithm (a) Residual Network for the Zero Flow; (b) Network after Augmenting Four Units along the P ath s 2 3 t ; (c) Network af ter Augmenting Four U nit along the P ath s 6 5 t In implementing any version of the generic augmenting path algorithm, we have the option of working directly on the original network with the flows or maintaining the residual network G( f ) and keeping track of the residual capacities and when the algorithm terminates recovering the actual flow variable To see how we can use
PAGE 19
11 either alternative, it is helpful to understand the relationship between edge flows in the original network and residual capacities in the residual network. T he concept of an augmenting path in the original network is considered next An augmenting path in the original network G is a path P (not neces sarily directed) from the source to the destination with on every forward edge ( i, j ) and on every backward edge ( i, j ) It is easy to show that the original network G contains an augmenting path with respect to a flow f if and only if the residual network G( f ) contains a directed path from the source to the destination Assuming we update the residual capacities at some point in the algorithm. What is the effect on the edge flows by units on edge (i j ) in the residual network corresponds to (1) an increase in by units in the original network, or (2) a decrease in by units in the original network, or (3) a convex combination of (1) and (2). We use the example given in Figure 2.4 (a) and the correspondin g residual network in Figure 2.4 (b) to il lustrate these possibilities. Augmenting four unit s of flow on the path s 6 2 3 t in the network produces the residual network in Figure 2.4 (c) with the corresponding edge flows shown in Figure 2.4 (d). Comp aring the solution in Figure 2.4 (d) with that in f igure 2.4 (a), we find that the flow augmentation increases the flow on edge (1, 2), (2, 4), (3, 5), (5, 6) and decreases the flow on edge (3, 4). Finally, suppose that we are given values for the residual capacities. How should we determine the flow ? Observe that since many combinations of and correspond to the same value of We can determine one such choice as follows. To highlight this choice, we rewrite
PAGE 20
12 as Now, if and ; otherwise, we set and (a) (b) (c) (d) Figure 2.4 Th e E ffect of Augm entation on Flow De composition (a ) Original Network with a Flow f ; (b) Residual Network for flow f ; (c) Res idual Network after Augmenting Four Unit s along the Path s 6 2 3 t ; (d) Flow in the Original Network after the Augment ation
PAGE 21
13 (a) (b) Figure 2.5 Flow Decomposition of the Solution in (a) Figu re 2.4(a) and (b) Figure 2.4(d) ( c ) ( d ) Figure 2.6 The Effect of Augmentation on F l ow Decomposition. (a) The Two A u gmentations P1 P2 P3 and Q1 Q2 Q3 ; (b) Net Effect of These Augmentations To obtain better insight concerning the augmenting path algorithm, we illustrate the effect of an augmentation on the flow decomposition on the preceding example. Figure 2.4 (a) gives the decomposition of the initial flow and Figure 2.4 ( d ) gives the dec omposition of the flow after we have augmented four unit s of flow on the path s 6 2
PAGE 22
14 3 t Although we augmented 4 unit s of flow along the path s 6 2 3 t the flow decomposition contains no such path. The path s 2 6 5 t defining the flow in Figure 2.4 (a) contains four segments: the path up to node 2 edge ( 2 6 ) and edge (6, 5) as a forward edge, and the path up to node t We can view this path as an augmentation on the zero flow. Similarly, the path s 6 2 3 t contains four segments: the path up to node 6 edge ( 6 2 ) as a backward edge, edge (2, 3) as a forward edge and the path down to node t We can view the augmentation on the path s 2 3 t as linking the initial segment of the path s 2 6 5 t with the last segment of the augmentation s 6 2 3 t linking the last segment of the path s 2 6 5 t with the initial segment of the augmentation s 6 2 3 t and canceling the flow on edge ( 2 6 ), which then drops from both the path s 2 6 5 t and the augmentation s 6 2 3 t In general, we can view each augmentation as pasting together segments of the current flow decomposition to obtain a new flow decomposition [ 5 ] Property 2 1 A flow f is a maximum flow if and only if the residual network G( f ) contains no a ugmenting path. Augmenting path algorithm and max flow min cut theorem are solutions for max flow problem with single source and single destin ation. They are the fundamental of the subgradient algorithm for network coding in Chap ter 5 T he following is t he pseudo code of the algorithm we used for the max flow min cut problem.
PAGE 23
15 B egin f := 0 (flow 0 on all edges) opt : = false W hile not opt do C onstruct the residual graph G(f) F ind a directed path P from S to T in G(f) I f such an augmenting path P exists T hen update flow f along P E lse set opt := true ; and S := the set of vertices in G(f) reachable from S E nd while R eturn f as the max flow, and ( S V S ) as the min cut E nd [8] Algorithm 2.2
PAGE 24
16 CHAPTER 3 MULTICAST CAPACITY The single multicast session from a source to a set of destinations is first considered It can be theoretically shown that achieving optimal throughput via multiple multicast trees is equivalent to the problem of Steiner tree packing which seeks to find the maximum number of pair wise edge disjoint Steiner trees, in each of which the multicast group remains connected. An intuitive explanation to such equivalence is that, each unit throughput corresponds to a unit information flow being transmitted along a tree that connects every node in the group. The maximum number of trees we can find corresponds to the optimal throughput for the session. Unfortunately, Steiner tree p acking problem has been shown to be NP complete. Several sub optimal algorithms have been developed to get the maximum throughput and the best ratio to the optimal throughput is 1/1.55 [ 6 ]. While we consider multiple multicast sessions, the problem gets more c omplicated. Each multicast session in isolation and independently may cause congestion on some links and reduce network utilization. Any participant can be a multicast source. As a r esult, a multicast distribution routing graph that connects multicast memb ers is shared by them. The choice of a multicast routing graph has several significant impacts both on the protocol performance and network utilization. In multicast packing problem, the choice of routing graph is more important since network resources nee d to be shared.
PAGE 25
17 There are two proposals for multicast routing backbone: tree based and ring based. A comparison of the optimal multicast tree and ring topology reports that closing the cycle may require as many as 25% more links. Thus, we focus on the pack ing of multicast trees. The underlying problem for the optimum shared tree is the Steiner tree proble m which is NP h ard as we introduced before. One a pproach is based on finding a median or core node and building a shortest path tree rooted at the core ( e.g., C ore B ased T rees (C B T) and Protocol Independent Multicast (PIM)). There are many variants of CBT approach depending on the definition of the core Furthermore, results provide a comparison basis between a single shared tree and multiple source spec ific trees. [9] 3.1 Group Communications Gro up communication refers to many to many communication sessions where multiple sources multicast independent data to the same group of receivers The a lgorithm for group communications with routing is proposed here F irst the maximum rate spanning trees are found based on Prim s algorithm. Then, those edges to non destination nodes are pruned This idea is demonstrated in [2] We will include additional s teps in to solve the group communications problem. E very multicast session can ha ve the same priority and share the network resources fairly. 3.2 Prim s Algorithm weighted graph It finds a subset of the edges that forms a tree that includes every vertex,
PAGE 26
18 where total weight of all the edges in the tree is minimized (or maximized). The algorithm continuously increases the size of a tree starting with a single vertex until it spans all the vertices. Input: A connected weighted graph with vertices V and edges E ; Initialize: Vnew = {x} where x is a node from V Enew = {}; Repeat until Vnew = V Choose edge ( u, v ) from E with minimal (maximal) weight such that u is in Vnew and v is not as shown in Figure 3; Add v to Vnew add (u,v) to Enew ; Output: Vnew and Enew describe the minimal (maximal) spanning tree. [2] Algorithm 3.1 Figure 3. 1
PAGE 27
19 (a) (b) (c) (d) (e ) Figure 3. 2 Ex
PAGE 28
20 Figure 3. 2 spanning tree from the source node 1 to all other nodes in the network. First, as shown in F igure 3. 2 (a), the algorithm compares two links 1 2 an d 1 6 from the reached node 1 to unreached nodes. Link 1 6 with the larger capacity has been chosen. Then, node 6 becomes a reached node. As shown in F igure 3. 2 (b), the algorithm compares the links 1 2, 6 2 and 6 5 from reached node 1 and 6 to unreached no des. Link 6 5 with the largest capacity has been chosen. Then, node 5 becomes a reached node. The algorithm keeps running in the same way until all nodes become reached nodes. 3.3 Algorithm for Group Communications gre edy algorithm for the group communications is as following. Step 1: algorithm. Prune the branches to the non destination nodes; Step 2 : Every tree decreases the bandwidth of the link s by ; Step 3 : If the bandwidth of every links are greater than zero, go to step 1 ; If the bandwidth of at least one link is less than 0 undo step 2 Go to step 4; Step 4 : If is greater than w and i = i+1 go to step 1 ; Else, stop; Step 5: Paste all the trees that every node has generated together. Algorithm 3.2 Here, is the bandwidth reduction step size, i is the times that step 4 has been run, the initial i = 0 m > 1 is the step size re duction factor, and is the
PAGE 29
21 stop threshold. The result will be better but the calculation will be more complex if decreases, m increases or w decreases. First of all, in order to get a fair result for each multicast session in group com munications, every session run each step of the algorithm at the same time. We first spanning tree prunes those links to non destination nodes. So, after the step 1 each session has a tree can cover from the source to all the destinations. Then, each tree decreases the capacity of the links it has by a same number After this step (step 2), each session has reserved a same number of capacity of the network, which means each session can transmit the same number of data with the reserved capacity of the network. Then, the algorithm reruns step 1 and step 2 unti l any of the remained bandwidth is less than zero. In step 4 if the stepsize is greater than the stop threshold w we cut the stepsize to and rerun the step 1 and step 2 This is because if the stepsize is too large, there wil l be too much waste of the capacity which will never be reserved and used by the group communications. If the stepsize is less than the stop threshold w the whole algorithm stops. After the whole algorithm stops, every session has its own distribu tion trees and all the sessions have the same maximum distribution rate s
PAGE 30
22 (a) (b) (c) (d) Figure 3. 3 Example of our Algorithm for Group Communications Figure 3. 3 shows an example of our algorithm for group commu nications. Figure 3. 3 (a) shows the topology of the network. We suppose all links are bidirectional here. So, we divide them into two links with opposite direction s In Figure 3. 3 (b), both sessions (session 1 with the source s1 and destinations d1 and d2 a nd session 2 with the source s2 and destinations d1 and d2) generate their own maximum capacity spanning trees with Prim s a lgorithm. In Figure 3. 3 (c), the algorithm prunes the un useful leaves of the trees. So, after the step 1, session 1 has its distribu tion tree s1 d1 s2 d2 and session 2 has
PAGE 31
23 its distribution tree s2 d1 and s2 d2 as shown in Figure 3. 3 (c). Then, each tree reserve s the same amount of capacity of the network which is 0.5 here in the example. Figure 3. 3 (d) shows the remained capacity of the network after step 2 So, with the reserved capacity, the group communications can guarantee 0.5 units/sec throughput for both sessions from source 1 and source 2 to all the destinations. For every round of the step 1 and step 2 the algorithm reserves som e capacity from the network. After the algorithm stop, each session can combine the trees with the reserved capacity together and get a routing scheme. This routing scheme can guarantee each session has the same throughput.
PAGE 32
24 CHAPT ER 4 MULTICAST ROUTING PROTOCOLS Internet Protocol ( IP ) network can transmit many types of data such as document files, audio and video within devices. However, when the data is sent via unicast, the source has to give each destination a copy which ma kes the whole network very inefficient. For example, when the manager wants to send his speech to the whole company, he has to send out one data flow for each employee. Obviously, this will cost a huge amount of the network resources including valuable Wi de Area Network ( WAN ) capacity. By using IP multicast technology, we can send data to a group of destinations through a very efficient way. Data flow is sent out from the source and tries to reach as far as possible in the network. Devices only copy the data when it has to send the data out through more than one interface in order to help the data reach all its destinations. 4.1 Compariso n between Multicast and Unicast When using unicast, the source needs to send out many copies of the data, each copy for each destination as shown in Figure 4.1.
PAGE 33
25 Figure 4.1 Example of Unicast When using multicast, the source sends out data through one data flow as shown in Figure 4.2. Figure 4.2 Example of Multicast A dvantages Increase the effic iency: Because the multicast do es not transmit several flows for each session, the network capacity has been used more efficiently. Optimize the performance: Multicasts avoids data flow redundancy. Data needs to be forwarded and proceeded becomes less. Support Distributed Application: When the topology increases, distributed has made a lot of new applications available such as Webcasting, Web TV,
PAGE 34
26 distance learning, teleme dicine, Web radio, real time videoconferencing, and other bandwidth and time critical information services 4.1.2 Most of the multicast applications use UDP. Comparing to unicast with TCP, it has several disadvantages. UDP may cause multicast group loss. So, multicast application has to count the unreliable factor into consideration. UDP does not have congestion control function. So, if UDP becomes more and more popular on n etwork, the network will become more congest and the whole performance of the network will drop. When the network topology changes, there is possibility that redundant multicast group will appear. There is potential security risk because the unwanted listener may find a way to join the multicast group. So, when design ing the multicast application, we have to take these d isadvantages into consideration. 4.2 Multicast Models Any Source Multicast (ASM), Source Filtered M ulticast (SFM), and S ource S pecific M ulticast (SSM) are three multicast models base d on how the receivers treat the multicast sources.
PAGE 35
27 4.2.1 ASM M odel In the ASM model, any sender can be a multicast source and send information to a multicast group. R eceivers can join a multicast group which is identified by the group IP address and o btain multicast i nformation being sent to the group address In this model, receivers can join or leave the multicast group at any time without regarding the sources 4.2.2 SFM M odel Not all multicast sources are valid in the SFM model. The upper layer s oftware checks the source address of received multicast packets and makes decision to permit or deny the traffic from specific sources Therefore, the receivers can only receive the data from part of the sources because some sources may be filtered 4.2.3 SSM M odel In the practical life, users may be interested in the multicast data from only some specific multicast sources. The receivers can specify their interested multicast sources in t he SSM model. 4.3 IGMP IGMP is the current widely used protocol for hosts to join or leave a specific should be forwarded to the hosts based on the IGMP request it received from the hosts. There are three versions of IGMP.
PAGE 36
28 4.3.1 IGMPv1 IGMPv1 is defined in RFC 1112 (Host Extensions for IP Multicasting). Figure 4.3 Example of IGMPv1 Assume that Host B and Host C are expected to receive multicast data to multicast group G1 with its group IP address. Host A is expected to receive multicast data to mul ticast group G2, as shown in F igure 4.3 The following shows how the hosts join the multicast groups and how the IGMP qu erier (Router B in Figure 4.3 ) maintains the multicast group memberships: The hosts send unsolicited IGMP reports to their interested addresses without having to wait for the IGMP queries from the IGMP querier.
PAGE 37
29 The IGMP querier periodically multicasts IGMP queries (with the destination address of 224.0.0.1) to all hosts and routers on the local subnet. Upon rece iving a query message, Host B or Host C (the delay timer of whichever expires first) announces its membership to G1 by sending an IGMP report to t he G1 multicast group address At the same time, Host A sends a report to the multicast group address of G2 The IGMP routers learn the memberships of G1 and G2 attached to the local subnet through the above query/report process. The multicast routing protocol like PIM generates (*, G1) and (*, G2) multicast forwarding entries on the router where represent s any multicast source. When the multicast data addressed to G1 or G2 reaches the IGMP router, the router forwards the data to lo cal subnets according to the multicast forwarding entries (*, G1) and (*, G2). IGMPv1 does not support Leave Group message for hosts to leave the multicast group. Hosts stop sending IGMPv1 report to the IGMP router whenever it wants to leave the multicast group. So, the IGMP router will delete the multicast forwarding entries for one multicast group only after a period of time without receiving any IGMPv1 report from hosts regarding that multicast group. 4.3.2 IGMPv2 Compared with IGMPv1, IGMPv2 has introduced a querier election mechanism and a L eave G roup mechanism.
PAGE 38
30 4.3.2.1 Que rier Election Mechanism In IGMPv1, the Des inated Router ( DR ) elected by the Layer 3 multicast routing protocol (such as PIM) serves as the querier among multiple routers on the same subnet. In IGMPv2, an independent querier election mechanism is introduced. The querier election process is as fo llows: Initially, every IGMPv2 router assumes itself as the querier and sends IGM P general query messages to all hosts and routers on the local subnet (the destination address is 224.0.0.1). E very IGMPv2 router compares the source IP address of the q uery message with its own interface address upon receiving a general query After comparison, the router with the lowest IP address wins the querier election and all other IGMPv2 routers become non queriers. All the non queriers start a timer, known as router receives an IGMP query from the querier before the timer expires, it resets this timer; otherwise, it assumes the querier to have timed out and initiates a new querier election process. 4.3.2.2 G M ech anism In IGMPv1, when a host leaves a multicast group, it does not send any notification to the multicast router. The multicast router relies on timeout of the host responding time to know whether a group no longer has members. This makes the leave grou p latency larger
PAGE 39
31 In IGMPv2, on the other hand, when a host leaves a multicast group: A Leave Group message (often referred to as leave message) is sent by the host to all routers (the destination address is 224.0.0.2) on the local subnet. Upon receiving the leave message, the querier sends a configurable number of group specific queries to the group being left. The destination address field and group address field of the message are both filled with the address of the multicast group being queri ed. One of the remaining members, if any on the subnet, of the group being queried should send a membership report within the maximum response time set in the query messages. If the querier receives a membership report for the group within the maximu m response time, it will maintain the memberships of the group; otherwise, the querier will assume that no hosts on the subnet are still interested in multicast traffic to that group and will stop maintaining the memberships of the group.
PAGE 40
32 4.3.3 IGMPv3 Figure 4.4 Example of IGMPv3 Built upon and being compatible with IGMPv1 and IGMPv2, IGMPv3 provides hosts with enhanced control capabilities and provides enhancements of query and report messages. IGMPv3 Enhances control capability of hosts. IGMPv3 h as introduced source filtering modes (Include and Exclude), so that a host not only can join a designated multicast group but also can specify to receive or reject multicast data from a designated multicast source. When a host joins a multicast group: a report with the Filter report w ith the Filter As shown in Figure 4.4 the network comprises two multicast sources, Source 1
PAGE 41
33 (S1) and Source 2 (S2), both of which can send multicast data to multicast group G. Host B is interested only i n the multicast data that Source 1 sends to G but not in the data from Source 2. In the case of IGMPv1 or IGMPv2, Host B cannot select multicast sources when it joins multicast group G. Therefore, multicast streams from both Source 1 and Source 2 will f low to Host B whether it needs them or not. When IGMPv3 is running between the hosts and routers, Host B can explicitly express its interest in the multicast data Source 1 sends to multicast group G (denoted as (S1, G)), rather than the multicast data S ource 2 sends to multicast group G (denoted as (S2, G)). Thus, only multicast data from Source 1 will be delivered to Host B [10] Currently, IGMPv2 is the most widely used protocol for hosts to joining the multicast group. 4.4 PIM Overview Protocol I ndependent Multicast (PIM) provides IP multicast forwarding by leveraging static routes or unicast routing tables generated by any unicast routing protocol, such as routing information protocol (RIP), open shortest path first (OSPF), intermediate system to intermediate system (IS IS), or border gateway protocol (BGP). Independent of the unicast routing protocols running on the device, multicast routing can be implemented as long as the corresponding multicast routing entries are created throug h unicast rout es. PIM uses the R eve rse P ath F orwarding (RPF) mechanism to implement multicast forwarding.
PAGE 42
34 Based on the implementation mechanism, PIM falls into two modes: Protocol Independent Multicast Dense Mode (PIM DM); Protocol Independent Multicast Sparse Mode (PIM SM). 4.4.1 PIM DM PIM multicast forwar ding, and is suitable for small size networks with densely distributed multicast members. PIM DM assumes that at least one multi cast group member exists on each subnet of a network, and therefore multicast data is flooded to all nodes on the network. Then, branches without multicast forwarding are pruned from the forwarding tree, leaving only those branches that contain receivers. periodically, that is, pruned branches resume multicast forwarding when the pruned state times out and then data is re flooded down these branches, and then are pruned again. When a new receiver on a previousl y pruned branch joins a multicast group, to reduce the join latency, PIM DM uses a graft mechanism to resume data forwarding to that branch. Generally speaking, the multicast forwarding path is a source tree, namely a forwarding tree with the multicast The tree is also called S hortest P ath T ree (SPT). PIM distribution. Although an SPT has the shortest path, it is b uilt with a low efficiency. Therefore the PIM DM mode is not suitable for large and medium size networks.
PAGE 43
35 4.4.2 PIM SM PIM multicast forwar ding, and is suitable for large and m edium size networks with sparsely and widely distributed multicast group members. PIM SM assumes that no hosts need to receive multicast data. In the PIM SM mode, routers must specifically request a particular multicast stream before the data is forward ed to them. The core task for PIM SM to implement multicast forwa rding is to build and maintain R endezvous P oint T rees (RPTs). An RPT is rooted at a router in the PIM domain as the common node, or R endezvous P oint (RP), through which the multicast data tra vels along the RPT and reaches the receivers. When a receiver is interested in the multicast data addressed to a specific multicast group, the router connected to this receiver sends a join message to the RP corresponding to that multicast group. The pat h along which the message goes hop by hop to the RP forms a branch of the RPT. When a multicast source sends multicast streams to a mu lticast group, the source side D esignated R outer (DR) first registers the multicast source with the RP by sending regis ter messages to the RP by unicast until it receives a register stop message from the RP. The arrival of a register message at the RP triggers the establishment of an SPT. Then, the multicast source sends subsequent multicast packets along the SPT to the RP Upon reaching the RP, the multicast packet is duplicated and delivered to the receivers along the RPT [10].
PAGE 44
36 4.5 Simulation of Current Multicast Protocols The current multicast protocols were simulated using QualN et [16] version 4.5 with a random ly gen erated 30 nodes wired network topology Link capacities are randomly picked between 10 M bps and 20 M bps. All link delays have been set to 0. Input and output buffer for all nodes is 16 Kbytes which is default. Packet size has been set to 1518 bytes/packet because in general larger packet size can get larger throughput in practice. Multicast with PIM DM and IGMPv2 has been enabled in all nodes as well as the whole environment. The unicast routing protocol is RIP which is good for small network that we simula ted here. All destination nodes join the multicast group from 0 s to 30 s of the simulation The data in group communications starts transmit from 10 s to 30 s to make sure that all multicast groups have been set up before the data transmission As packe t size has been set up, the interval time between each packet was modified to adjust the throughput with binary search to get the maximum throughput for group communications. No packet loss is allowed and throughput for each so urce and each destination mus t be the same.
PAGE 45
37 Figure 4.5 Network Topology for the Simulation of Current IP Multicast Protocols Figure 4.5 shows the topology of the network. As the simulated protocol does not have the function to adjust the capacity in undirected network s we set bi directional links for all connected nodes. The maximum throughput from source 5 and 15 to destinations 1 and 2 is 8.1 M bps; the maximum throughput from source 5 and 15 to all other 28 nodes as destinations is 4.8 M bps; the maximum thr oughput from source 1, 2, 3, 4 and 5 to destinations 6 and 7 is 2.7 M bps; the maximum thr oughput from source 1, 2, 3, 4 and 5 to all other 25 nodes as destinations is 2.5 M bps; the maximum throughput from sour ce 1 to 10 to destinations 11 and 12 is 2.2 M bps; the maximum throu ghput from 1 to 10 to all
PAGE 46
38 other 20 nodes as destinations is 1.4 M bps. (All source and destination numbers mentioned here are the node number s in the network topology). The simulation result will be compared with routing and network coding algorithms s imulation results later in Chapter 6.
PAGE 47
39 CHAPTER 5 NETWORK CODING Network coding is a recent topic in information theory that allows the nodes to generate output data by encoding its received data. Thus, nodes may mix the input packets together and send them out as fewer packets. Figure 5.1 Multicast with R outing Figure 5. 1 shows the multicast routing mechanism. We assume the network is a delay free and error free network. S is the source while Y and Z are the sinks. All the links are with unit capacity. As shown in the graph, each sink could receive 3 units in 2 seconds. So, the maximum throughput for this multicast application is 1.5 units/sec here by using routing. The bound here is the cut T Y, W X and U Z is shared by two sinks. So, the maximum throughput is 3/2 which is 1.5 units/sec.
PAGE 48
40 Figure 5. 2 Multicast with Network C oding Figure 5. 2 shows the multicast with network coding mechanism. The net work condition and the application re quirement are both the same as F igure 5. 1. While we allow node W encode its two input packets into one packet, the throughput becomes to 2 units/sec here. Node Y can receive two independent packets b1 and at the same time. Y can decode these two ind ependent packets to get both and Also, Z can get both and by independent packets and The bound here is the minimum min cut for each sink which decide how many independent packets a sink could receive. 5.1 Linear Network Coding The output flow at a given node is obtained as linear combination of its input flows when we use linear network coding [11 ]. Linear network coding is sufficient to guarantee the multicast throughput as the same as the maximum throughput from the source to each individual destination in unicast session When the packet s to be combined have diffe rent size s the shorter ones are padded with trailing 0s. Assume that each packet consists of L bits. We can interpret s consecutive bits of a packe t as a symbol over the finite field with each packet consisting of a vector of L/s symbols. With linear
PAGE 49
41 network coding, outgoing packets are linear combinations of the original packets, where addition and multiplication are performed over the finite field Figure 5. 3 Network Coding [1 2 ] Following are the two equations for intermediate nod es to encode the packets Eq. 5.1 Eq. 5.2 is the local encoding vector. Each intermediate node has it s own local encoding vector which can be both ass igned by an algorithm or random generated. The intermediate node uses its local encoding vector to combine the input data together and send them out. Following two equations are global view of the linear network coding E q. 5.3 Eq. 5.4 Eq. 5.5
PAGE 50
42 where is the global encoding vector. The global encoding vector is calculated by local encoding vectors. With the global en coding vector, the destination nodes can decode the received packets to those symbols from the source. Following are the equations for decoding Eq. 5.6 Eq. 5.7 Each destination node t can recover the source symbols as long as the matrix formed by the global encoding vectors, has full rank n 5.2 Practical Random Network Coding There are several ways to generate the local encoding vector. In practice, distributed algorithm is preferred because it may be difficult to get global knowledge of the whole network. So, practical ra ndom network codi ng has been introduced by Chou [ 13 ] When using practical random network coding, intermediate nodes s elect the linear coefficients in a finite field of opportune size in a random way The encoding vector is included within the encoded pa cket. Nodes store within their buffers the received packets This a llows asynchronous packets arrivals and departures with arbitra rily varying rates, delay and loss Simulation results indicate that even for small field sizes (for example, s = 8) the proba bility of selecting linearly dependent combinations becomes negligible.
PAGE 51
43 When a node receives a packet, it decides whether to store the packet or discard it If the new packet increases the current rank of the matrix, it is an innovative packet. If it do es not increase the rank of the matrix, it means that the packet contains redundant information and it is not needed for decoding the original source packets. Hence, non innovative packets ar e dropped Generally, we hope to invert G t by collect ing n or mo re packets However, we can use the early decoding mechanism which is recommended here. Nodes p erform Ga ussian elimination after receiving each packet Every node detects and discards non innovative packets G t tends to be lower triangular, so it is typica lly possible to decode x 1 x k with fewer more than k packets This can make m uch shorter decoding delay than block decoding It has been shown that, in a directed network with network coding scheme, a multicast rate is feasible if and only if it is feas ible for a unicast from the sender to each receiver [1] Thus, there is an explicit max flow min cut capacity bound for the single source multicas t network coding problem Also, the research work proves that linear coding usually suffices in achieving this maximum rate [ 11 ] There exist directed graphs where the throughput gains of network coding for multicasting can be very significant. However, in undirected graphs the throughput gain is at most a f actor of two [3] Experimental results over the network graphs of six Internet service provides showed a small throughput gain in this case [ 9 ]. Li et al. introduced a c flow algorithm using linear programming to find the maximum throughput for one multicast session in undirected network s using network coding [3] Then, they introduced a sub gradient algorithm with less complexity and
PAGE 52
44 distribu ted implementation. They also gave the c flow algorithm for multiple multicast sessions which they called m flow algorithm. However, the m flow algorithm cannot guarantee the same throughput for each source in group communications and it does not allow int er session coding Li et al. showed that, the throughput advantage with network coding is always 1.0 for one multicast session in 1,000 randomly generated undirected benefi t of network coding is not higher optimal throughput, but to facilitate significantly more efficient computation and implementation of strategies to achieve such optimal 5. 3 Subgradient Algorithm Step 1: Choose initial orientation (e.g., balanced orientation) Step 2: Repeat Compute S Ti max flow, for all i Refine orientation: Increase bandwidth share for saturated links Decrease bandwidth share for under utiliz ed links Until convergence As a result, optimal orientation obtained Step 3: Compute S Ti max flow, for all i As a result, optimal multicast rate and routing strategy obtained Step 4: Randomized code assignment As a result, complete transmission strategy obtained [2]. Algorithm 5.1
PAGE 53
45 Primal variables in the orientation c are updated in two steps. First, we compute a new orientation vector as follows: Eq. 5.8 w here is a prescribed sequence of step s izes. A simple sequence is for some positive constants a, b and c. is valued to 1 when the link is the min cut or 0 when the link is not the min cut. Eq. 5.9 w here We can get the final after it converges. During each iteration of orientation refinement, the algorithm computes the max flow/min cut from the sender to each receiver, and increases the capacity shared for more saturated links, while decreases th e capacity shared for under utilized links. 5. 4 Algorithm for Network Coding in Group Communications Group communication refers to many to many communication sessions where multiple sources multicast independent data to the same group of receivers. Rese archers always treat group communication as a simple problem by adding a super source which is connected to all the sources with unbounded capacity links. However, it is not able to control the fairness between different sources in this method. Take the fo llowing network as an example.
PAGE 54
46 Figure 5. 4 Group C ommunication s with N etwork C oding In the Figure 5. 4 T and U are the sources while Y and Z are the sinks. If we add a super source S and connect it to T and U with unbounded capacity link, we can get the maximum multicast throughput from S to Y and Z with network coding is 5 units/sec based on the max flow min cut bound. This result comes from the cut T Y and X Y or the cut U Z and X Z. However, the maximum sending rate for the source T is only 1 units/se c because of the cut T W and Y X and the maximum sending rate for the source U is 4 units/sec because of the cut U W and Z X. The two independent sources are not fair here. In the example of Figure 5. 4 the network is a directed network. If we extend t he discussion to undirected network s we can find that different sources may compete on the link capacity. As a result, we cannot treat group communication s as a simple one multicast session problem by just adding a super source.
PAGE 55
47 Now, let s take a lo ok at another example to show that the method of adding only a super source and considering the group communication with network coding as a single session multicast problem is even incorrect. Figure 5. 5 Six Nodes Bi directional Network In the network shown in Figure 5. 5 node 1 and 3 are the sources while node 2, 4 and 5 are the sinks. The minimum min cut from the super source to the three destinations is cut 1 which means the maximum sending rate of the super source is 19. The minimum min cut from th e source node 1 to the three destinations is cut 2 which means the maximum sending rate of the source node 1 is 8. The minimum min cut from the source node 3 to the three destinations is cut 3 which means the maximum sending rate of the source node 3 is 8. The interesting thing is that 19 is larger the 8+8. As we know, the max flow min cut theory shows that the sink node 4 will never be able to receive more than 8 independent units per second from the source node 1 because of the cut 2 whenever we use netwo rk coding or not. The same thing happens from the source node 3
PAGE 56
48 to the sinks node 2 and 5 because of the cut 3. As a result, in the original network (without the super source), the source 1 and source 3 will never be able to send out information with a tot al sending rate that is larger than 16. So, the result we get from the method that adding a super source is incorrect We need additional steps to design this new subgradient algorithm Step 1: Choose initial orientation (e.g., balanced orientation) Step 2: Put the super source S in and connect it to all sources with unbounded capacity links Step 3: Repeat Compute Si Tj max flow, for all i and j Find the minimum max flow fsingle for all sources to all sinks Compute S Tj max flow for all j Fi nd the minimum max flow fsuper for S to all sinks Refine orientation: Increase bandwidth share for saturated links Decrease bandwidth share for under utilized links Until convergence As a result, optimal orientation obtained Step 4: Compute S Tj max flow fsuper for all j Compare this with n times the minimum max flow for all sources and sinks fsingle and choose the less one. As a result, optimal multicast rate and routing strategy obtained Step 5: Randomized code assignment As a resul t, complete transmission strategy obtained Algorithm 5.2
PAGE 57
49 Primal variables in the orientation c are updated in two steps. First, we compute a new orientation vector as follows: Eq. 5.10 If fsuper <= n*fsingle when belongs to the min cut of the super source If fsuper > n*fsingle when belongs to the min cut for fsingle Where is a prescribed sequence of step sizes and i indicates the ith source. A simple sequence is for some positive constants a, b and c is the maximum flow for the source i. Eq. 5.11 Where We can get the final after it converges. When there are n indepe ndent sources in the network, the main concern is trying to maximize the capacity from the super source to the destinations. However, when the maximum flow for the super source comes more than n times the maximum flow for any single sources, we have to bal ance the capacity for those sources. During each iteration of orientation refinement, the algorithm computes the max flow/min cut from each sender (include the super source) to each receiver. After that, it increases the capacity shared for more saturated links, while decreases the capacity shared for under utilized links. This method will maximize the capacity from the sources to the sinks and make different
PAGE 58
50 sources as fair as possible. After getting the final if the fsuper is less than n times fsingle we can run the well designed random coding scheme by treating this as a one session multicast from the super source to all the receivers while each sources sends the information with rate fsuper/n If the fsuper is larger than n times fsingle we use the same random coding scheme while each source can only sends the information with rate fsingle
PAGE 59
51 CHAPTER 6 SIMULATION AND COMPARISON OF RESULTS T he main differences b etween the s ub gradient a lgorithm for n etwork c oding and the greedy algorithm for r outing in group communications are as follows. The algorithm for routing tries to help each session (transmission from each source) reserve some capacity from the network in each iteration until there is no m ore available capacity. There is unused capacity in the network. However, it requires more computation. So, we can modify the parameter of the algorithm based on the network scenario. The algorithm for network coding is based on the idea that network cod ing can guarantee each session (transmission from each source) reaching the throughput obtained by max flow min cut algorithm to each destination. So, the algorithm assigns the half and half capacity on each direction of every link in the network initially Then, different sessions compete for the capacity of the network until it converges. The main contribution of these algorithms is that they can both guarantee each sessions having the same throughput. Figure 6.1 shows the network topology for simulat ion The topology is randomly generated by the n etwork simulation tool Qual N et version 4.5 Thirty nodes are randomly connected through links with random integer capacities between 10 and 20 units/sec.
PAGE 60
52 Figure 6.1 Simulation Network Topology Both of the algorithms were simulated in MATLAB with the following network parameters Routing: i nitial s tepsize 0.1, b ound 0.1 ; Network c oding: k <= 1000, s tepsize 1000/ (50+k) k is the iteration number.
PAGE 61
53 (a) (b) (c) Figure 6.2 Simulation Results of the Comparison of Netwo rk Coding and Routing (a) Group Communication with 2 Sources; (b) Group Communications with 5 Sources; (c) Group Communications with 10 Sources
PAGE 62
54 Figure 6.2 (a) shows the result of group communications with 2 sources which are node 5 and node 15 here F igure 6.2 (b) shows the result of group communications with 5 sources which are node 1, node 2, node 3, node 4 and node 5 here Figure 6.2 (c) shows the result of group communications with 10 sources which are node 1, node 2, node 3, node 4, node 5, node 6 node 7, node 8, node 9 and node 10 here. As shown in Figure 6.2, n etwork coding do es have obvious throughput advantage sometimes, but not always. We changed the topology, the source nodes as well as the destination nodes for the simulation and r esults were always similar. At most of the time, network coding does not have obvious throughput advantage which is the same as one multicast session communications s imulated by Li et al. [2]. The following results are for comparison with current IP multicat p rotocols. Figure 6.3 Maximum Throughput s from Source Node 5 and 15
PAGE 63
55 Figure 6.4 Maximum Throughput s from Source Node 1, 2, 3, 4 and 5 Figure 6.5 Maximum Throughput s from Source Node 1 to Node 10
PAGE 64
56 So, our simulation showed a throughput compariso n between network coding and routing for group communicatio ns in undirected network s Current multicast protocols are much more scalable and easier to deploy in real network. However, the throughput with those protocols will be much less than the algorithm s we simulated here. Figure 6.3 to Figure 6.5 show that both algorithms being introduced here (one for routing only and one for network coding) have huge throughput benefit to current IP multicast protocols. Also, although we can set up QOS policy in route rs and switches, current protocols cannot optimize the network to make different sessions as fair as possible. Trade off is always there in engineering problems. Researchers are doing huge effort on network coding and hope that network coding can support t he maximum throughput with routing and becomes easier to deploy
PAGE 65
57 CHAPTER 7 CONCLUSION AND FUTURE WORK 7.1 Summary The traditional method to solve the group communications problem is putting a super source with unlimited bandwidth to all s ources. In this thesis, it is shown that this method cannot guarantee the fairness within different sources for routing. Also, the method can be incorrect in certain scenarios Two algorithms are presented in this thesis one for routing and one for networ k coding to guarantee that each source has the same fairness and get the sub optimal throughput for group communications in undirected network s All current widely used routing protocols are topology based. The throughputs using both these algorithms (one for routing only and one for network coding ) are much better than current widely used IP multicast protocols. Between the two proposed algorithms the algorithm for network coding can have throughput benefit in some scenarios but not always. Here, we show that network coding does not have constant throughput benefit in undirected network s in group communications scenario with the consideration of fairness within different sources. 7.2 Recommendations for Future Research Both our algorithms are not distri buted algorism s. So, it is hard to deploy them into large networks. Also, they will not work well if there are link failures or topology
PAGE 66
58 change s in the network. Further research is necessary to improve the scalability and robustness of these algorithms. Cl uster based algorithms might be a good approach. We only simulated these algorithms in one randomly generated thirty nodes network scenario. Simulation in large network is necessary for future research.
PAGE 67
59 REFERENCES [ 1] R. Ahlswede, N. Cai, S. Li, and R. W. Yeung, "Network Information F low," IEEE Trans. on Information Theory, vo l. 46, p p 1204 1216, 2000 [2] Y. W u, P. A. Chou, and K. Jain, "A Comparison of N etwo rk Cod ing and Tree P acking," in Information Theor y, ISI T 2004 Proceedings p p 143 149, 2004 [3] Z. Li, B. Li D. Jiang, and L. C. Lau, "On Achieving Optimal Throughput with Network C oding," in Proceed ings IEEE INFOCOM 2005 vol. 3, pp. 2184 2194, 2005 [4] R. Dougherty, Chris Freiling, and Kenneth Zeger, "Un achievability of Network Coding Capacity," IEEE Trans. on Information Theory, v ol. 52 p p 2365 2372, June 2006 [ 5 ] J. B. Orlin, T. L. Magnanti, and R. K. Ahuja, Network Flows: Theory, Algorithms, and A pplications 1993 [ 6 ] G. Robins and A. Zelikovsky, "Improved Steiner Tree Approximation in G raphs," in Proceedings of the Eleventh A nnu al ACM SIAM S ymposium on Discrete A lgorithms p p 770 779, 2000 [ 7 ] U. S. R. Murty and J. A. Bondy, Graph Theory with A pplications ," 1976 [8] A. Mirzaian, http://www.cse.yorku.ca/~aaw/Wang/MaxFlowMinCutAlg.html [9] S. Chen, O. Gunluk, and B. Yener, The Multicast Packing Problem, IEEE/ACM Trans. Networking, p p 311 318, June 2000 [10] Hangzhou H 3C Technologies Co., Limited, Document Team, www.h3c.com [ 11 ] S. Li, R. W. Yeung, and N. Cai, "Linear N etwork C oding," IEEE Trans. on Information Theory, vol. 49, p p 371 381, 2003 [12] E. Fasolo Slides from www.cs.virgin ia.edu/~yw5s/Network%20coding.ppt [13] P.A. Chou, T. Wu, and K. J ai n, "Practical N etwork C oding," the 51 st Allerton Conf. Communication, Control and Computing, Oct. 2003
PAGE 68
60 [ 14] R. Koetter, and M. Medard, "An A gebr aic A pproach to N etwork C oding," IEEE/ACM Tra n. on Networking, vol. 11, p p 782 795, Nov. 2003 [ 1 5 ] N. J.A. Harvey, R. Kleinberg, and A.R. Lehman, On the Capacity of Information Networks, IEEE Trans. Information Theory, pp. 2345 2364, June 2006 [16] Scalable Network Technologies, Inc., http://www.scalable networks.com/ [17] L. R. F ord, and Journal of Mathematics 8, pp. 399 404, 1956
