Chapter 1 Introduction 1.1 Background Recently wireless mobile ad-hoc networks (MANETs) have received much attention from the research community for their important applications in emergence and military situations, and other areas. Most current multi-hop wireless mobile ad-hoc network implementations suffer from severe throughput limitations, less robustness and do not scale well. Typically, MANETs are resource constrained nodes, e.g. battery power can be limited, and therefore higher data throughput cannot always be achieved by increasing transmission power. Overall data throughout for MANETs is also an issue as communications are often multi-hop in nature. It is therefore a challenging research problem to send more information with low power to optimize the throughput. To improve network throughput, a novel idea of network coding for the current packet switched networks has been proposed by Ahlswede et al. [1]. Traditionally, the intermediate node in the network just forwards the input packets to the intended nodes. However, network coding allows the intermediate node to combine some input packets into one or several output packets based on the assumption that the intended nodes are able to decode these combined packets. Figure 1 is a simple illustration for this promising idea which shows how network coding can save a great deal of transmissions, and thus improving the overall wireless network throughput. In the three-node scenario in Figure 1, Alice and Bob want to send packets to each other with the help of a relay. Without network coding, Alice sends packet p1 to the relay 1 and then the relay sends it to Bob. Likewise, Bob sends packet p2 to Alice. Therefore, a total of 4 transmissions are required for Alice and Bob to exchange one pair of packets. With network coding, the relay combines packets p1 and p2 together simply using XOR and then broadcasts it to both Alice and Bob. Alice and Bob then extract the packets they want by performing the required decoding operation. The total number of transmissions is therefore reduced to 3. This illustrates the basic idea of how network coding is able to improve the network throughput. (a) No coding (b) Coding Figure 1: A simplified illustration of network coding, showing how network coding saves bandwidth consumption. It shows Alice and Bob to exchange a pair of packets using 3 transmissions instead of 4. Recently the research focus on network coding has shifted towards the practical aspects of research, in particular, for wireless mesh networks. For example, COPE [2] is regarded as the first practical implementation of network coding in wireless mesh networks. In [2-5] the authors introduce COPE as a new packet forwarding architecture which combines several packets together by bit-wise exclusive OR (XOR) operation, 2 coupled with a completely opportunistic approach in routing packets. COPE inserts a coding shim between the IP and MAC layers, which identifies coding opportunities and benefits from them by forwarding multiple packets in a single transmission. What is more, by taking advantage of the opportunistic property and simple XOR coding algorithm, COPE manages to address practical issues when integrating network coding in the current communication protocol stack. The details of the opportunistic property and practical considerations will be described in the next chapter. Experimental results [2-5] have shown that COPE can substantially improve the throughput of wireless mesh networks by 3 to 4 folds in simple one-hop topologies and it also slightly improves the throughput in large scale multi-hop wireless mesh networks. 1.2 Motivations Though S. Katti [2] has shown COPE to work well in wireless mesh networks, there are still a number of issues to be further investigated. For example, it is shown that in a large scale multi-hop wireless environment, COPE is unable to deal with TCP traffic as well as UDP traffic. This was because the retransmission scheme and congestion control in TCP protocol would degrade the performance. In general, network coding causes packet reordering, and for TCP traffic, the performance suffers further. On the other hand, it is observed that even for UDP traffic, the performance of COPE in large scale networks is also much worse than that of single-hop topologies. One of the main reasons could be that the impact of COPE’s control messages on the overall performance of the opportunistic network coding scheme has not been studied. It is further suspected that if COPE’s control messages are scheduled intelligently, the performance of the opportunistic network coding scheme in large scale scenario may be improved. 3 In addition, few works have been reported on designing a practical network coding scheme for wireless mobile ad-hoc networks. MANETs play a very important role in many fields of application nowadays, for example the emergency situation and military application. However, the throughput and robustness of wireless mobile ad-hoc networks are limited due to many factors, such as the node resource and changing topologies. We are therefore interested in investigating the outstanding issues described here with the aim to improve the performance of opportunistic network coding in wireless mobile ad-hoc networks. In this thesis we propose to first study and investigate the performance behavior and key control parameters of opportunistic network coding in large scale static wireless mesh networks. Our proposed approach is to be independent of routing protocol so that this opportunistic network coding scheme can take advantage of any new designed or existing routing protocols. To validate our proposed schemes, we choose QualNet as the pseudo implementation environment and simulation platform since QualNet is the most realistic network simulator there is, and its protocols are fully implemented similar to that of actual implementation. So our proposed solutions and simulation results would be much closer to real system implementations. Due to the time restriction of this project, my experiment only deals with UDP traffic rather than those various traffic in the real world and are not able to be implemented in large scale mobile ad-hoc network scenarios. However, the valuable simulation results and analysis presented in this thesis would shed light on further research about the complex traffic situation within mobile environment. 4 1.3 Thesis Contributions This thesis has carried out the following work and contributed to the understanding of performance of practical network coding in wireless mesh networks. In addition, detailed studies on the behavior and key control parameters of COPE on simple topologies have shed light on the way to address the issue on large scale wireless networks. Lastly, based on these valuable insights, we have proposed and designed an intelligent opportunistic network coding scheme which would be particularly suitable for the MANET environment. We summarize our contributions as follows: 1. Extend the QualNet simulator to include opportunistic network coding functionalities – We have designed and developed the new functionalities in the existing protocol stack of QualNet. We further integrate these new functionalities into QualNet as a new enhanced network layer protocol. Through this implementation on QualNet, the performance of any kind of opportunistic network coding scheme and network size can be easily studied through QualNet simulations. 2. Evaluate and study the behavior of opportunistic network coding scheme via our enhanced QualNet both on simple as well as large scale topologies – We enhance the original COPE functionality and simulate its behavior in Alice-and-Bob (i.e. 1to-1), X and Cross topologies. Based on the simulation results from these simple topologies we further study its performance in a 20-node multi-hop wireless mesh network. 3. Evaluate and study the key parameters of opportunistic network coding scheme with respect to overall network throughput, e.g. the impact of control message on the performance of opportunistic network coding scheme. We then use the findings 5 for further study of network coding scheme on large scale wireless ad-hoc networks. We find out there is an optimal value for the control message interval to maximally improve the network throughput. 4. We propose an intelligent opportunistic network coding scheme that is suitable for large scale wireless ad-hoc networks and demonstrated the effectiveness of our solutions through simulations. The proposed intelligent algorithm manages to reduce the overhead and interference caused by control message in large scale multi-hop networks without degrading the benefit brought by network coding. 1.4 Related Works Network coding is regarded as a promising technique to improve network throughput. It originates from Ahlswede et al [1], which demonstrates that intermediate nodes in network may combine several received packets into one or several output packets. Much theoretical work has been done to optimize network coding scheme in information and networking systems. Li et al. [6] has showed that linear codes are sufficient for multicast traffic to achieve the maximum capacity bounds. At the same time, Koetter and Medard [7] has proposed an algebraic approach and showed that coding and decoding can be done in polynomial time. Ho et al. [8] has presented the concept of random linear network coding, which makes network coding more practical, especially in distributed networks such as wireless networks. In paper [4] an intra-flow network coding scheme is proposed to deal with intra-flow traffic, which can effectively handle reliability issue. Joon-Sang [9] has presented a network coding based ad-hoc multicast protocol CodeCast, which is especially well suited for multimedia application in a wireless network. All of the above works have shown results either analytically and/or through extensive simulations. 6 In the last few years, many researchers have focused on developing practical network coding techniques in wireless networks [10-12] for inter-flow traffic, to significantly improve the network capacity. A great deal of attention has been focused on dealing with the practical issues and designing the implementable protocols with network coding [13-15]. Generally, the common practical issues are that how to integrate network coding technique into the current network protocol stacks and achieve low complexity of coding and decoding scheme. In [2] S. Katti et al proposed COPE which is regarded as the first practical network coding implementation for wireless mesh network dealing with inter-flow and unicast traffic. With the opportunistic listening and opportunistic coding characteristics, COPE exploits the broadcast nature of the wireless channel. Through eavesdropping and sharing information among neighbors, the intermediate node with COPE can simply XOR multiple packets into one packet and broadcast it to several neighbors. All the neighbors are then able to decode the specific packets from the combined packet by the simple XOR method. The authors show that this opportunistic network coding scheme improves network throughput several times for wireless mesh network. In addition, K. Ajoy et al [21] further evaluated the performance of COPE with two different routing protocols, i.e. AODV and OLSR using three different queue management schemes such as FIFO, RED and RIO. The authors show that OLSR provides better performance than AODV for COPE while the FIFO achieves shortest packet delay among the three queue management schemes. Finally in paper [16] which is a valuable primer for network coding the authors explicitly explain some popular network coding schemes and also describe the advantages and challenges in research area and practical implementation. They also enumerate some other promising fields where network coding 7 could be applied, from peer-to-peer (P2P) file distribution networks to wireless ad-hoc networks, from the improvement of network capacity to increase of network security. 1.5 Thesis Organization The rest of this thesis is organized as follows. Chapter 2 presents the overview of opportunistic network coding including two components of the opportunistic character, i.e. opportunistic listening and opportunistic coding. We also explain the packets coding and decoding algorithm implemented in this specific opportunistic network coding scheme. In chapter 3, we present the design of opportunistic network coding algorithm and architecture. A detailed description of the control flow is presented to show how we design this opportunistic network coding scheme. Chapter 4 shows our implementation of the opportunistic network coding scheme in QualNet simulator. We first introduce the architecture of QualNet simulator, including the protocol model and application program interface. Then we enumerate some key programming abstractions of our implementation in QualNet, which show how we implement the opportunistic network coding scheme and how we integrate them into QualNet simulator. In chapter 5, we present our simulation results and analyze the observations, including the parameters that affect the results and the way to improve the coding scheme. In chapter 6, an intelligent version of opportunistic network coding scheme is proposed based on the conclusions drawn in previous chapter. We further demonstrate how this intelligent scheme would be suitable for MANETs. 8 Chapter 7 presents the performance evaluation of the intelligent opportunistic network coding scheme proposed in previous chapter. In the last chapter, Chapter 8, we draw some conclusions based on those simulations. We also highlight a number of areas where further enhancements can be explored. 9 Chapter 2 Opportunistic Network Coding Overview COPE is a new forwarding architecture for current packet switched network especially for wireless mesh networks with unicast traffic. Traditionally, the intermediate node in the wireless network directly forwards the input packet to the next hop. However, in COPE the intermediate node may XOR several input packets into one output packet and then broadcast the combined packet to several intended neighbors based on the assumption that all the intended neighbors are able to decode the combined packet. No synchronization or prior knowledge of senders, receivers, or traffic rates are necessary i.e. all of which may vary at any time. COPE depends highly on the local information shared by neighbors to detect and exploit coding opportunities whenever they arise. To successfully exchange the information among neighbors in a wireless environment, an opportunistic mechanism is introduced which has two main components, i.e. 1. Opportunistic Listening 2. Opportunistic Coding In this chapter, all these techniques will be explained and corresponding coding and decoding schemes will be also introduced. 2.1 Opportunistic Listening The wireless network comprises a broadcast medium wherein the nodes in the network are able to hear packets even when they are not the intended recipients. This is the basic idea of opportunistic listening which allows the nodes to snoop in the network and store all overheard packets in their local buffers, called the packet pool. To achieve this, all 10 nodes in the network should be equipped with omni-direction antennas and set to “promiscuous” listening mode. A specific callback function is necessary to handle the packets heard in the promiscuous mode. We will introduce how the callback function is implemented in QualNet simulator in Chapter 4. The packet pool, in our implementation, is an FIFO queue with fixed capacity of N0. The larger the N0 is, the more packets the node can store, therefore, there would be more opportunities to perform network coding. However, if N0 is too large, the packets in the head of packet pool would be too old, thus degrading the coding efficiency. In our implementation, N0 is set at 128. The total amount of storage required is less than 192 kilobytes, which is easily available on today’s PCs, laptops or PDAs. This constitutes the Opportunistic Listening function. In addition to promiscuous listening, each node periodically sends reception reports to its neighbors to share information with them which is important to the coding decision. This reception report contains the information on the packets that the host has heard and stored in its packet pool. Once the neighbors have received these reception reports, they extract all the relevant information and store them in another local buffer, called the report pool. The reception reports are normally inserted into the data packets as an extra packet header and sent together with the data. However, when nodes have no data packets to send, they will periodically broadcast “Hello” messages which contain the reception reports to their neighbors. In the original COPE implementation, this Hello message is the same as the control packet, but in our implementation, control packets are indicated as a separate kind of packet. The Hello message would be part of an intelligent scheduling scheme, which will be described at Chapter 6. 11 In summary, through the Opportunistic Listening technique, the nodes learn and share their states with each other, thus contributing to the nodes’ network coding decision. 2.2 Opportunistic Coding Opportunistic coding allows nodes to combine multiple packets into a single packet based on the assumption that all intended recipients can extract the packets they want from the combined packet. However, the main issue is regarding which packets to code, and how to code. Each node should answer this question based on local information and without further consulting with other nodes. A basic-method to address the coding issue in wireless networks is that each node maintains a FIFO queue of packets to be forwarded. When the MAC indicates that the node can send, the node picks the packets at the head of the queue, checks which other packets in the queue which may be encoded with this packet, XORs those packets together, and broadcasts the single combined packet. However, the question is which packets should be combined together to maximize network throughput since a node may have multiple coding options. It should therefore pick the one that maximizes the number of packets delivered in a single transmission. In [2] the authors provide a good example to illustrate this situation. In Figure 2(b) node E has eight packets in its output queue P1-P8. The list in Figure 2(a) shows the next hop of each packet in E’s output queue. When the MAC notifies node E to transmit, node E dequeues packet P1 from head of the output queue and tries to code with other packets in the output queue. With the help of the information in reception reports, node E knows what packets its neighbors have in the Packet Pool. Now node E has some coding options as shown in Figure 2(c). 12 The first option is P1 ⊕ P2, which is not a good option because none of the recipient nodes can decode the P1 ⊕ P2. The second option in Figure 2(c) shows a better coding decision P1 ⊕ P3. As node C has packet P3 so it can successfully decode packet P1; and node D has packet P3 so it can decode packet P1 successfully. As for the third and fourth options, it can be seen that these two are bad coding decisions as none of the recipient nodes can decode either P1 ⊕ P3 ⊕ P5 or P1 ⊕ P3+ ⊕ P4. One of the interesting observations is that the second option is a better coding decision than the third and fourth options although the third and fourth options can code more packets than the second option. The fifth option P1 ⊕ P3 ⊕ P6 is a much better option than the second option as three of the recipient nodes (B, C and D) can decode their intended packet successfully, i.e. node B has P1 and P3 so it can decode P6, node C has P1 and P6 so it can decode P3 and node D has P3 and P6 so it can decode P1 successfully. Coding option sixth is the worst as it would have coded four packets and none of the intended next-hop can decode the encoded packet successfully. The seventh option P1 ⊕ P2 ⊕ P6 ⊕ P8 in the Figure 2(c) is the best coding decision for that scenario with the maximum number of packets XORed together. Here node A has P1, P3 and P6 so it can decode P8 successfully. Similarly node B, C and D can decode their intended packet P6, P3 and P1 respectively from the encoded packet of P1 ⊕ P2 ⊕ P6 ⊕ P8. As can be seen from this simple example, a general opportunistic coding rule can be like this proposed in paper [2]: A node can XOR n packets p1, … , pn together to transmit to n next-hops r1, … , rn only if each next-hop ri has all n-1 packets pj for j ≠ i. Each time when a node is ready to send, it tries to find the maximum n in order to code and transmit as many packets as possible in a single packet. This coding scheme has a few 13 other important characteristics. First, there is no scheduling or assumed synchronization. Second, no packet is delayed; every time the node sends a packet it picks the head of the queue as it would have done in the current approach. The difference is that, whenever possible, the node tries to load each transmission with additional information through coding. Third, the scheme does not cause packet reordering as it considers the packets according to their order in the FIFO queue, for both transmission and coding. This characteristic is particularly important for TCP flows, which may mistake packet reordering as a congestion signal. 2.3 Packet Coding Algorithm In this section, we introduce the details of packet coding algorithm based on the opportunistic coding idea mentioned above. Some practical issues in the implementation and the specific solutions are also proposed. First, the coding scheme in the original COPE does not introduce additional delay. The node always de-queues the head of its output queue and checks coding opportunities when it is ready to send. If there is no coding opportunity, the node will send the packet without waiting for the arrival of a matching codable packet. However, in our design, we have added a waiting scheme before the coding procedure shown in Figure 3. This waiting scheme takes action immediately before the network coding procedure. When the node is ready to send, it checks the packet in its output queue and executes this waiting scheme before it executes the network coding. This waiting scheme works as follows: a node checks the coding opportunity if and only if the number of packets in its output queue is greater than N and no new packet arrives during the last T seconds. N and T are called queue threshold and waiting duration of the waiting scheme, respectively. The values of N 14 P1 D P5 B P2 C P6 B P3 C P7 A P4 C P8 A (a) Next Hop of packets in E’s Queue A’s Packet Pool P3 P1 B’s Packet P6 P8 P1 P3 E’s Output Queue P1 P2 P3 P4 P5 P6 C’s Packet Pool P8 P1 P7 P8 D’s Packet Pool P6 P8 P3 P6 (b) Sending Scenario Coding Option Is it good? P1 P2 Bad P1 P3 Better P1 P3 P4 Bad P1 P3 P5 Bad Coding Option P1 P3 Is it good? P6 Much better P1 P3 P6 P7 Bad P1 P3 P6 P8 Best (c) Possible Coding Option Figure 2: An Example of Opportunistic Coding and T would significantly impact the performance of the algorithm. It is understandable that the waiting scheme would accumulate N packets in the output queue thus improving the coding opportunities with the price of delaying the packet by T. However, the 15 improvement in network throughput may shorten the packet delay which may compensate for the increase in delay due to the waiting scheme. Figure 3: A simplified illustration of waiting scheme, showing a threshold N in the output queue. A node sends packet if and only if the threshold N is fulfilled and no packets came during the last T seconds. Second, the coding scheme gives preference to XOR-ing packets of similar lengths, because XOR-ing small packets with larger one reduces overall bandwidth savings. Empirical studies show that the packet-size distribution in the Internet is bimodal with peaks at 40 and 1500 bytes [17]. We can therefore limit the overhead of searching for packets with the right sizes by distinguishing between small and large packets. We might still have to XOR packets of different sizes only when necessary. In this case, the shorter packets are padded with zeros and the receiving node can easily remove the padding by checking the packet-size field in the IP header of each native packet. Third, the coding scheme will never encode together packets headed to the same nexthop, since the nexthop will not be able to decode them. Hence, we only need to consider packets headed to different next hops. Hence the relay node maintains a virtual queue for each neighbor. When a new packet is inserted into the output queue, an entry is added into the virtual queue for the specific intended neighbor. 16 Finally, we want to ensure that each intended neighbor is able to decode its native packet from the combined packet. Thus, for each packet in output queue, our relay node checks whether each of its neighbors has already heard the packet. The neighbor’s information is shared and learned by the reception report, mentioned previously. In our implementation, each node maintains the following data structures. Each node has 3 FIFO queues of packets to be forwarded, which we call the output queue, which is the default node configuration. All these 3 queues have different priorities from 0 to 2. In our implementation, data packets have lowest priority 0 while Hello messages and control messages have the highest priority 2. For each neighbor, the node maintains two per-neighbor virtual queues, one for small packets (e.g. smaller than 100 bytes), and the other for large packets. The virtual queue for a neighbor A contain pointers to the packets in the output queue whose nexthop is A. Each node maintains two extra buffers named packet pool and reception report pool. Packet pool is used to store the overheard packets and reception report pool is used to store the reception report from neighbors. The packet pool stores the native packets while the reception report pool stores the report indicating the specific details of the packets overheard by neighbors. One entry of the report contains the packet’s ID and previous hop and next hop’s address. The details of the packet format will be explained in next Chapter. The specific coding procedure is illustrated by the following pseudo-code. 17 Coding procedure – Pick packet p at the head of the output queue. Sending Packets = {p} Nexthops = {nexthop (p)} if size (p)>100 bytes then queue = virtual queue for large packet Else queue = virtual queue for small packet End if for Neighbor i = 1 to M do Pick packet pi the head of virtual queue Q (i, queue) if ∀ n ∈ Nexthops ∪ {i}, n can decode p ⊕ pi based on the reception report p = p ⊕ pi Sending Packets = Sending Packets ∪ {pi} Nexthops = Nexthops ∪ {i} end if end for queue = ! queue for Neighbor i = 1 to M do Pick packet pi, the head of virtual queue Q (i, queue) if ∀ n ∈ Nexthops ∪ {i}, n can decode p ⊕ pi based on the reception report p = p ⊕ pi Sending Packets = Sending Packets ∪ {pi} Nexthops = Nexthops ∪ {i} end if end for return Sending Packets 18 In above pseudo-code p represents the specific packet in the output queue while Q indicates the whole output queue in term of the overall virtual queue structure for all neighbours. The variable queue is a two-state variable indicating which queue is selected between those two queues for large and small packets. In this algorithm, it uses “!” operation to switch between these two states. For example, queue=!queue means if originally queue indicates the one for large packets, after the operation, queue indicates the one for small packets and vice verse. Given the values of i and queue, Q(i, queue) can locate the specific virtual queue for neighbour i. In the original coding scheme, the authors introduced an intelligent guessing scheme which is based on the integrated ETX [18] routing scheme. At the congestion situation, the reception report may be dropped at the wireless channel or may be too late to reach the intended node. Thus the relay node may miss some coding opportunity. Depending on the packet delivery probability of the wireless link, the relay node may guess whether the intended neighbour has received the packets or not. Even though the authors show the intelligent guessing technique can somehow benefit the total network coding opportunity at congestion situation in static wireless mesh networks, we do not implement this technique in our design for several reasons. First, we would like to design an opportunistic network coding scheme independent of routing protocol, thus making the algorithm very flexible at difference scenarios by cooperating with the suitable routing protocol instead of integrating with ETX based routing algorithm. Furthermore, ETX algorithm calculates the metric value by measuring the loss rate of broadcast packets between a pair of neighbour nodes, indicating the link quality. However, this method is not suitable for the wireless mobile environment, because the topology is always changing 19 as well as the link quality. In contrast, the AODV [19] and OLSR [20] are two more practical routing protocols for wireless mobile ad-hoc networks than those based on ETX. 2.4 Packet Decoding The packet decoding scheme is much simpler than the coding side. As mentioned above, in COPE each node maintains an extra buffer, named the packet pool, to store a copy of the packets overheard or sent out. Each packet is indicated by a specific packet ID consisting of the packet source address and IP sequence number. When a node receives an encoded packet consisting of n native packets, the node goes through ids of the native packets one by one in the local packet pool and retrieves the corresponding n-1 packets. Ultimately it XORs these n-1 packets with the received encoded packet to get the intended packet. 2.5 Pseudo Broadcast In COPE, the node sends the packets in an encoded manner, which is a single encoded packet that contains information of several packets with several different next hops. Moreover, for opportunistic listening, all nodes snoop at the network to monitor all packets that are transmitted among its neighbors. The natural method to do this would be broadcast. However, one of the biggest disadvantages of 802.11 MAC protocol is that a recipient does not send an acknowledgement in response to a broadcast packet. In the absence of an acknowledgement, the broadcast mode offers no retransmissions and consequently very low reliability. In addition, a broadcast source does not detect collision, and thus does not back off and retransmit. If several nodes sharing the same wireless channel broadcast data packets to the neighbors, the total network throughput would be severely degraded due to the congestion. On the other hand, unicast mode ensures both 20 sending node retransmission and back-off but only to one specific destination at a time. However, unicast does not help in opportunistic listening and consequently not in opportunistic coding. To address this problem pseudo broadcast is introduced in COPE. Pseudo broadcast is actually unicast and therefore benefit from the reliability and the back-off mechanism. The link layer destination field of the encoded packet is set to the MAC address of one of the intended recipients. However, an extra header after the link layer header is added, listing all other next hops of the encoded packet (except link layer destination). Recall that all nodes in the network listen on promiscuous mode. They snoop in the network by eavesdropping all packets transmitted among neighbors. In this way the node is able to process those packets not headed to it. When a node hears an encoded packet, it checks the link layer destination field to determine if it is the intended receiver or not. If it is, it will process this packet directly. If not, this node further checks the next hop list in the next packet header to see whether it is the intended receiver or not. If not, it just stores a copy of that packet-as a native packet-in the packet pool. If it is meant for the next hop, it processes this encoded packet further to retrieve the intended packet and then stores a copy of the decoded native packet in its packet pool. As all packets are sent using 802.11 unicast, the MAC layer is able to detect collisions and back-off properly. Pseudo broadcast is therefore more reliable than simple broadcast and inherits all the advantages of broadcast. In this chapter, we have presented the overview of the opportunistic network coding for wireless environment. In next chapter, we will show the specific architecture of opportunistic network coding in our implementation including the specific packet header structure and the whole control flow of the algorithm. 21 Chapter 3 Opportunistic Network Coding Architecture In this chapter, we introduce the architecture of the opportunistic network coding scheme which we have implemented in QualNet simulator. The details of the architecture are based on the overview of the opportunistic characteristics, coding and decoding algorithm introduced in the previous chapter. Here the packet header structure will be shown and the functionality of each field in the header will be explained. Next the overall control flowchart will be presented to illustrate the structure of the opportunistic network coding algorithm in our implementation. 3.1 Packet header Figure 4 shows the modified variable-length coding header for the opportunistic network coding scheme, which is inserted into each packet. If the routing protocol has its own header, our coding header sits in between the routing and MAC layer headers. Otherwise, it sits between the IP and MAC headers. Only the shaded fields in Figure 4, is required in every coding header (called the constant block). Besides this, there are two other header blocks containing the identifiers of the coded native packets and the reception reports. Constant block: The first block records some constant values for the whole coding header. For example, it records the number of coded native packets in this encoded packet, the number of reception reports attached in this header, the packet sequence number and the total length of the header. Besides these, the protocol serial information and parity can also be inserted in this constant block. In our implementation, we have added the version information and check-sum fields in this block. 22 Figure 4: Packet header for our algorithm. The first constant block indicates the number of entries in the following blocks. The second block identifies the native packets encoded and their next hops. The last block contains reception reports. Each entry identifies a source, the last IP sequence number received from the source, and a 32-bit long bit-map of most recent packets seen from that source. Identifiers (Ids) of the coded native packets: This block records metadata to enable packet decoding. The number of entries is indicated in the first constant block. Each entry contains the information of corresponding native packet. It begins with the packet Id, which is a 32-bit hash of the packet’s source IP address and IP sequence number. This is followed by the IP address of the native packet’s nexthop. When a node hears an XOR-ed packet, it checks the list of nexthop in this block to see whether it is the intended nexthop of this XOR-ed packet, in which case it decodes the packet, and processes it further. Reception reports: As shown in Figure 4, reception reports form the last block in the header, and the number of report entries is also recorded in the first constant block. Each report entry specifies the source of the reported packet SRC_IP, which is followed by the IP sequence number of the last packet received from the source LAST_PKT, and a 23 bit-map of recently heard packets. This bit-map technique for reception report has two advantages: compactness and effectiveness. In particular, it allows the nodes to report each packet multiple times with minimal overhead. This prevents reception reports from being lost at high traffic congestion situation. Our packet header structure by-and-large follow the original COPE header structure; however, in order to make this opportunistic mechanism fit in QualNet and the mobile environment, we make some modifications in our implementation. First, we remove the asynchronous acknowledgment scheme in the original COPE. Originally, COPE exploits hop-by-hop ACKs and retransmission to guarantee the reliability in the hop-by-hop fashion, thus adding an ACKs block at the end of the header structure. However, we discover this asynchronous ACKs technique is not a good solution; sometime it even makes the performance worse. So in our header structure, there is no ACKs block, thus resulting in smaller overhead. Besides this, we rearrange the header structure thus placing a constant block at the beginning of the header shown in Figure 4. In addition, we use 32-bit bitmap instead of 8-bit at the reception report block. The bitmap is used to represent packets. For example, if the first bit of the bitmap indicates packet 10, then the second bit indicates packet 11 and so on. A longer bitmap can indicate more packets than a short one, which is used to compensate for the delay associated with the reception report. This is particularly so in a mobile environment, where a node can provide more information to a new incoming neighbor which has no information about his neighbor node, therefore potentially improve the coding opportunity. Another modification is that we replace the MAC address at the Nexthop field of the second block in the header with IP address. IPv4 address which is 16-bit shorter than MAC address 24 means less overhead in the header thus compensating for the long bitmap. Moreover, the replacement could make our solution independent of the underlying MAC layer, making it suitable for different networks. 3.2 Control Flow This section describes the overall packet flow of our opportunistic network coding scheme, which mainly consists of two parts, i.e. the Sending Side and the Receiving Side. 3.2.1 Sending Side (a) Sender Side (b) Receiver Side Figure 5: Flowcharts for our opportunistic network coding implementation 25 Figure 5(a) shows the flowchart of the Sender Side. When a node is ready to send a packet, it first checks whether it needs to wait for the next new input packet or not. Based on the waiting scheme we have introduced in previous chapter, the node just checks the number of packets in its output queue and whether it has received new packets in the last T seconds. If the number of packet in the output queue is less than the threshold value N or if the wait period of T seconds has yet to expire, the node will wait for additional new packets. Otherwise, the node goes to the next step immediately to de-queue from the head of the output queue. Next the node continues to traverse the packets in the output queue to pick up some packets that are able to be coded with the head packet according to the coding algorithm. After the packets are XOR-ed together, the node constructs the header block with the Ids of the coded native packets in the header followed by the reception reports. Finally, the combined packet with the extra coding header is transmitted. Alternatively, if no other packets can be coded with the head packet, the native head packet is just added with the coding header and transmitted without any delay. 3.2.2 Receiving Side At the receiver side, whenever a node receives a packet, it checks whether this packet is a coded packet or not. If this packet is native without the extra coding header, the node processes it in the usual way. Otherwise, the node processes it according to the flowchart of the receiving side illustrated in Figure 5(b). First, the node extracts the reception reports from the header and updates the neighbor’s state recorded in its report pool. Next, the node checks whether there is more than one packet combined together in this packet, in which case the node tries to decode it. After the node manages to get the native packet, it stores a copy of the native packet in its packet pool and goes on to check 26 whether it is the intended next hop. If not, it just stops handling this packet. Otherwise, it passes this native packet to higher level protocol for further processing. In the next chapter, we will describe our implementation details of this opportunistic network coding architecture in QualNet simulator. 27 Chapter 4 Implementation in QualNet Simulator In this chapter, we describe the details of our implementation of opportunistic network coding scheme in QualNet simulator. Before describing the detailed programming, it is useful to present a brief description of the QualNet simulator. 4.1 Simulator Abstraction The QualNet simulator is a commercial network simulation tool derived from GloMoSim that was first released at 2000 by Scalable Network Technologies (SNT). We use QualNet 4.0 as our performance evaluation study platform. QualNet is famous for its ultra high-fidelity, scalability, portability, and extensibility. It can simulate a scenario with thousands of network nodes because it takes full advantage of multi-threading capability of multi-core 64 bit processor. In addition, the source code and configuration files are identical to those in the real communication system organized in OSI protocol stack model. Furthermore, the protocol architecture in QualNet is also very close to the real TCP/IP network structure, which consists of the Application, Transport, Network, Link (MAC) and Physical Layers, from top to bottom. Compared to other open source network simulation tools, QualNet is closest to the real system implementation, thus capable generating more realistic and accurate simulation results. 4.1.1 Discrete event simulator Compared to the continuous time simulator, a discrete event simulator is much more popular among the fields of industry implementation or academic research. There are many articles online providing many convincing evidence and analytics on this 28 statement, so we do not discuss further in this thesis. QualNet is also a discrete event simulator where the system state changes over time only when an event occurs. An event can be anything in the network system such as a packet generation request, a collision or time out and so on, which triggers the system to change its state or perform a specific operation. In QualNet there are two event types: Packet events and Timer events. Packet events are used to simulate exchange of data packets between layers or nodes. To send a packet to the adjacent layer, the QualNet kernel passes the handle to the specific node and the node schedules a packet event for the adjacent layer, then returns the handle back to kernel. After a pre-set delay, the occurrence of the packet event simulates the arrival of the data packet, which triggers the QualNet kernel passing the handle to this node again and the adjacent layer in this node processes the data packet further. Next, this data packet is passed to another adjacent layer until it is freed. Packet events are also used for modeling communication between different nodes in the network. In fact, the communication among nodes in the network can only be achieved by scheduling data packets and exchanging them with each other. On the other hand, Timer events are used to perform the function of alarms. For example, a time alarm is used to trigger the periodic broadcast of control message every one second. Timer events are very useful and important for the simulator to schedule more complex event pattern like in a real application system. In QualNet, both the Packet event and Timer event are defined via the same message data structure. A message contains the information about an event such as the event type, sequence number, generating node and the associated data. Figure 6 shows the message data structure in QualNet. 29 struct message_str { Message* next; // For kernel use only. short layerType; // Layer which will receive the message short protocolType; // Protocol which will receive the message in the layer. short instanceId; short eventType; // Message's Event type. …… MessageInfoHeader infoArray[MAX_INFO_FIELDS]; int packetSize; char *packet; NodeAddress originatingNodeId; int sequenceNumber; int originatingProtocol; int numberOfHeaders; int headerProtocols[MAX_HEADERS]; int headerSizes[MAX_HEADERS]; Figure 6: Message structure in QualNet Some of the fields of message data structure are explained bellow. layerType: Layer associated with event indicating which layer will receive this message protocolType: Protocol associated with event indicating which protocol will process this message in the layer instanceId: For multiple instances of a protocol, this field indicates which instance will receive this message 30 infoArray: Stores additional information that used in the processing of events and the information that needs to be transported between layers. Packet: This field is optional. If the event is for actual data packet, this field would hold the data. Headers added by different layer are included in this field. packetSize: The total size of the packet field. numberOfHeaders: Recording how many headers are added headerProtocols: It is an array storing the specific protocol which adds the header. headerSizes: It is an array storing the specific size for each header. The last three fields are for packet tracing. If the packet tracing function is enabled, these fields are filled up during the simulation to facilitate the analysis after the simulation, but it slows down the simulation. All the fields listed above are some key parts in the message data structure, more details can be found in the API reference provided by QualNet. We have added some new entries in the message structure to facilitate our implementation of network coding scheme, which will be described later. 4.1.2 Protocol Model in QualNet As mentioned above, each node in QualNet runs a protocol stack just like the physical communication device in the real world and each protocol operates at one of the layers in the stack. Before we implement our own protocol into QualNet, we describe how a protocol is modeled in QualNet. Figure 7 shows the general protocol model as a finite state machine in QualNet. 31 Figure 7: Packet Model in QualNet A general protocol model in QualNet consists of three states, Initialization, Event Dispatcher and Finalization. In the Initialization state, the protocol reads parameters from the simulation configuration file to configure its initial state. Then the protocol transfers to the Event Dispatcher state, which is the kernel of the protocol model. This state contains two sub-states, Wait For Event state and Event Handler state, which construct a loop. Initially, all protocols are waiting for the events headed to them to happen, in which case the protocol transfers to the Event Handler state to call the specific handler function to process the event. Afterwards, the protocol transfers back to the Wait For Event state to wait for the next event to happen. After all potential events in the simulation are processed, the protocol transfers to the last state called Finalization, in which the protocol may print out the packet tracing data and also the simulation statistics into some output files. 32 4.1.3 Application Program Interface QualNet provides several Application Program Interface (API) functions for the events operations. Some of these APIs can be called from any layer while others can only be called by specific layers. The complete list of APIs and their explicit descriptions can be found in the API reference provided by QualNet. Here we select some examples which are helpful for our implementation in QualNet. MESSAGE_Alloc: This API allocates a new Message structure, which is called when a new message has to be sent through the system. MESSAGE_Free: It is called to free a message when this message is no longer needed in the system. MESSAGE_AddInfo: This is to allocate one “info” field with given info type for the message. MESSAGE_RemoveInfo: Remove one “info” field with given info type from the info array of the message. MESSAGE_PacketAlloc: Allocate the “payLoad” field for the packet to be delivered. It is called when a message is a packet event carrying specific data. MESSAGE_AddHeader: Add a new header with the specific header size to the packet enclosed in the message. MESSAGE_RemoveHeader: Remove the header from the packet enclosed in the message. MESSAGE_Send: It is called to pass message within QualNet. 33 IO_ReadNodeInput: It is called to read the parameters from external configuration file. IO_PrintStat: It is called to print out the statistics collected during the simulation into an output file. NetworkIpSneakPeekAtMacPacket: Called directly by the MAC layer, this allows a routing protocol to “sneak a peek” or “tap” messages it would not normally see from the MAC layer. Besides these APIs to process the message, there are many other APIs enclosed in Scheduler and Queue classes that deal with the queuing system. For example, some APIs are used to insert a packet into or de-queue a packet from a queue while others are used to construct some kinds of queuing systems. There are also some APIs for queue management, such as FIFO queue, RIO queue or RED queue. 4.1.4 QualNet Simulator Architecture From the section of protocol model in QualNet, we have learned a protocol is modeled as a finite state machine with three states: Initialization, Event Dispatcher and Finalization. However, how are those protocols managed as a stack in QualNet similar to the TCP/IP protocol stack in the real world? As we know the protocols are grouped into layers in the protocol stack of TCP/IP model, which is achieved by registering those protocols into the event type table and protocol type table managed by QualNet. What is more, the corresponding event handler function should be embedded into each layer’s entrance. Take the AODV routing protocol as an example. In a wireless ad-hoc network with the nodes equipped with AODV routing protocol, when a packet is passed to the 34 network layer from the transport layer, the entrance function in the network layer is checked to determine whether this packet needs to be routed or not. If it needs to be routed, the network layer entrance function calls the embedded Event Handler function of AODV protocol registered at protocol table to process this packet further. Another issue is how this protocol stack operates in QualNet. As mentioned above, a protocol model in QualNet has three components: Initialization, Event Dispatcher and Finalization, which operate in hierarchical manner, first at the node level, second at the layer level and finally at the protocol level. At the start of the simulation, each node in the network is initialized by the kernel of QualNet. The initialization function of the node calls the initialize function for layer initialization. The layers are initialized in a bottom up order. All the layers are initialized one node at a time, except the MAC layer which is initialized locally. Each layer initialization function then calls all the protocol initialization functions running in that layer. The initialization functions of a protocol create and initialize the protocol state variables. If the value of the variable is not given by the user, a default value is selected during the Initialize state. After all the nodes are initialized, the simulator is ready to generate events and process events. When an event occurs, the QualNet kernel passes the handle to the node where the event occurs. The node calls a dispatcher function to determine which layer should process the event further. The event dispatcher function of that selected layer then calls the event dispatcher function for the appropriate protocol based on the protocol type information enclosed in the event, normally in the packet header. The protocol event 35 dispatcher function then calls the corresponding event handler function to perform the actions for the occurred event. At the end of the simulation, Finalization functions are called automatically and hierarchically in a manner similar to Initialization functions. Finalization functions usually print out the statistics collected during the simulation time. 4.2 Programming Abstraction in QualNet As we know a native QualNet does not support any network coding functionality, we should therefore implement the opportunistic network coding functionality in QualNet’s protocol stack before we can further evaluate its performance via simulations. In this section we will describe the implementation details of our opportunistic network coding scheme in QualNet. As mentioned in chapter 2, this opportunistic network coding scheme works between the MAC and IP layers by inserting an extra coding header between MAC and IP headers. However, in order to be consistent with the real network architecture, we do not create an extra independent communication layer between Link and Network layer. Instead, we implement the opportunistic network coding protocol in the network layer but at the bottom entrance. It means the opportunistic network coding protocol will process the packet passed from MAC layer first before passing it to the IP protocol. On the other hand, before the IP protocol passes the IP packet to the MAC layer, the network coding protocol will traverse the packets in the output queue to search for the opportunity of combining several packets together. Before describing the details of the functions, we will present the states and variables that our network coding protocol maintains. Figure 8 shows the overall map of structures created for the protocol. 36 37 Figure 8: Implemented data structure The main structure maintained by the protocol is the CopeData which contains other small part data structures. It is initialized by the specific protocol initialization function, CopeInit(), at the beginning of the simulation. Recall the protocol model in QualNet, where the protocol initialization function is called hierarchically to read the parameters from the external configuration files to initialize the protocol states. In addition, as can be seen there are two local buffers in the CopeData, i.e. “ppScheduler” and “reportInfo”. “ppScheduler” is managed as a FIFO queue and is the packet pool storing the overheard packet from neighbors. There are three functions to manage this queue, one of which is CopePpQueueInit() to initialize this queue system, and another is CopePpQueueInsert() to insert one cope of the new overheard packet into the local packet pool. The last is CopePpQueueExtract() to extract specific packets from this packet pool for decoding. The capacity of this queue is set to the default value of QualNet and when the queue is full, and the head of the queue will be dropped automatically due to the characteristic of FIFO. The other local buffer is the “reportInfo”, which, in fact, is a hash table to keep the reception report information for the neighbors and itself. The head of this table is the self reception report records. When a new neighbor sends a reception report to this node, CreateSelfReport() function is called to insert a new element in the “reportInfo” hash table. If an existing neighbor sends new information to this node, CopeReportUpdate() and Cope_SubReportUpdate() functions is called to update the records for this neighbor. To facilitate debug, we add another function CopePrintReportInfo() to print out all the content of this local buffer. The length of this hash table is not limited since the number of neighbors for one specific node would not be too large to consume large amount of memory. However, the specific array for each 38 element in this hash table is upper bounded by MAX_Entry, which is set at 128 in our implementation. Besides these two local buffers, the protocol also keeps a structure, CopeStats, to collect the statistics during the simulation. These statistics are printed out by CopeFinalize() function at the end of simulation. Another important variable is pkt_seq_no which records the local packet IP sequence number to help identify a specific IP packet with the packet original source address in the network. Some other entries in the CopeData structure are variables to store the parameter values configured in the external configuration file, such as “helloInterval” storing the Hello message interval, “processHello” storing the Boolean value indicating whether to process Hello message and so on. Besides the main data structure CopeData for the protocol to maintain the protocol states, CopeHeaderType structure is used by the protocol to indicate the extra coding header structure. It is created according to the packet header structure discussed in chapter 3. The functions CopeAddHeader() and CopeRemoveHeader() are used to add or remove the extra coding header into or away from the IP packet. Having explained the overall data structures required by this opportunistic network coding scheme, we next describe our programming details of the algorithm. For the sender side, the challenging part is checking the coding opportunity and intelligently combining the packets together. Based on the flowchart shown in Figure 5(a) in Chapter 3, we construct a more detailed modeling diagram for the sender side function CopeCoding() shown in Figure 9. 39 40 Figure 9: CopeCoding() flowchart This function acts immediately after an IP packet is successfully de-queued from the output queue of network layer. First, it updates the self reception report and inserts a copy of this de-queued packet into the local packet pool. Then it initializes the coding header structure waiting for the corresponding components to be filled up later. The next stage is to traverse the packets in the output queue to check whether there is any packet that can be combined with the de-queued packet. The checking and coding functionalities are implemented in Coding() function. There are two criteria to check the coding opportunity. The first is that no more than one packet is heading to the same recipient; the other is that each of the intended recipients has received all other packets except the one destined for it. The Coding() function achieves the second criteria by checking the reception reports stored in the local report pool. After all the codable packets in the output queue are found, the Coding() function combines them together via a simple XOR method then generates the list of encoded packet’ ids which is inserted into the coding header. After the coding procedure, a report is generated and sent to the neighbors. The report contains the new packets’ information received recently by this node, which will be attached at the bottom of coding header. Finally, the CopeAddHeader() function is called to assemble all the header’s components together and add it into the encoded packet which is passed to the MAC layer later. We have so far discussed the implementation details of the sender side, Figure 10 shows the flowchart on the receiver side function, CopeSneakPeekAtMacPacket(). 41 Figure 10: CopeSneakPeekAtMacPacket() flowchart 42 As mentioned at Chapter 2 all nodes in the network work in promiscuous mode and the opportunistic network coding scheme takes advantage of pseudo broadcast, so the node is able to process the packet even if it is not the intended recipient. This is achieved by calling MAC_SneakPeekAtMacPacket() function at MAC layer. This function is provided as a default API at MAC layer to enable promiscuous mode. Further this function calls the NetworkIpSneakPeekAtMacPacket() function implemented at network layer to allow network CopeSneakPeekAtMacPacket() layer is to process implemented this as packet. the child Our function function of NetworkIpSneakPeekAtMacPacket(), thus managing to make the network coding scheme sitting on the network layer to process the encoded packets. The first step of our function is to check whether the packet is carrying the coding header. If not, the packet is just dropped. Otherwise, the CopeRemoveHeader() function is called to remove the coding header from this packet, and meanwhile the coding information and reception report enclosed are extracted. The coding information is used to help the decoding procedure while the reception report for the neighbor is used to update the neighbor’s state recorded in local report pool. Next, the node goes through the list of nexthop enclosed in the second block of the header to check whether it is one of the next hops. If it is not and there is only one packet enclosed, the node updates its reception report and store one copy of this native packet in the packet pool. Otherwise, the packet will be dropped. On the other hand, if this node is one of the next hops, the node takes advantage of the coding information extracted from the header to decode the native packet headed to it, and then updates its report and stores one copy of the packet in the local packet pool. After that, the network coding protocol passes the native IP packet to the IP protocol for further process. 43 We have presented our implementation details in QualNet simulator in this chapter, the performance evaluation will be discussed in next chapter. 44 Chapter 5 Performance Evaluation and Discussion 5.1 Emulation Environment We have implemented the opportunistic network coding algorithm in QualNet 4.0 simulator as described in the previous chapter. In this chapter we will evaluate and analyze the performance of this opportunistic network coding scheme over several simple topologies in a wireless mesh network setting. In our simulation, we use CBR (Constant Bit Rate) as the application traffic, which is carried on UDP transport protocol. CBR is the simplest traffic pattern to model the streaming internet traffic, such as audio or video traffic. CBR is chosen to simplify the analysis of our simulation results. More complex application traffic will be studied in the future. For the CBR application, the packet size is set at 512 bytes while the packet interval varies from 0.1s to 0.01s, which means the actual load for each source node is from 41kbps to 410kbps. Each source node starts its transmission at the first second while the entire simulation lasts 2 minutes. 5.1.1 Simulation Parameters In this section we present the details of the simulation configuration. Our simulation uses IPv4 (Internet Protocol Version 4) as the network layer protocol with three output queues with different priorities while the Internet Control Message Protocol (ICMP) model is disabled. IEEE 802.11 is chosen as the MAC layer protocol and the propagation delay is set at 1µs. Both RTS and CTS mechanisms are enabled in our simulation with RTS threshold value set at zero. The wireless channel frequency is set at 2.4GHz. The signal propagation model is statistical with a limit of -110.0dBm. Two Ray 45 ground model is used to model the propagation pass loss. Shadowing is modeled as constant with mean set at 4.0 dB. Channel fading effect is not considered. The noise factor of the channel is 10.0. IEEE 802.11b is selected as the radio type with nominal data rate of 2 Mb/sec. The transceiver is equipped with omni-directional antenna with an efficiency of 0.8. Antenna mismatch and antenna connection loss is set at 0.3 dB and 0.2 dB, respectively. The transmission power of the antenna is 15.0 mWatt. For 2Mb/sec data rate the receiver sensitivity is set to -89.0. The estimated directional antenna gain is 15, and the packet reception is modeled as Phy802.11b. Besides these pre-determined parameters for the device, there are also some parameters for the opportunistic network coding algorithm. For the waiting scheme, the waiting duration T is set at 40 milliseconds. Further in-depth study of the impact of T is a part of the future work. The queue threshold of waiting scheme N increases from 0 to 37 which is a bit larger than the default setting of the output queue in QualNet simulator thus large enough to study the behavior of the waiting scheme. The Hello message interval T0 is initially set at 40 milliseconds; however, it will be adjusted in our further study later. 5.1.2 Experimental Topology First we run our simulation on three simple network topologies, Alice and Bob, X and Cross topologies. We then run each topology on a 20-node multi-hop scenario. Aliceand-Bob scenario has been shown in Figure 1. Figure 11(a) shows the X topology where node 1, node 4 and relay can hear each other and so do node 2, node 5 and relay. However, node 1 and node 4 cannot hear node 2 and node 5, and vice verse. In this scenario node 1 and node 2 send packets to node 5 and node 4 respectively under the help of the relay node. The Cross topology is shown in Figure 11(b) where there are 5 nodes placed as a 46 cross. The relay can be heard by all other nodes, and each of other nodes can hear 2 nearest nodes beside the relay. How many nodes can be heard by any one node in the network is an important consideration. The neighborhood condition and local traffic pattern are the essential differences between Cross and X topologies. In the Cross scenario, 4 traffic flows travel cross the relay. Node 1 and node 5 communicate with each other and so do node 2 and node 4. Figure 12 shows the topology of the 20-node scenario. There are (a) X topology (b) Cross topology Figure 11: X and Cross wireless topologies Figure 12: 20-node multi-hop topology, where all nodes uniformly placed on 1000m*1000m area 47 20 nodes uniformly distributed on 1000m*1000 m area. Node 1 and node 5 communicate with each other and AODV is used as the routing protocol. 5.1.3 Performance Metrics In our simulation, we use three metrics to evaluate the performance of our opportunistic network coding scheme by comparing with the situation without network coding. The three comparison metrics are total packets dropped, network throughput and average end-to-end packet delay. The total packets dropped metric denotes the total number of packets dropped at the end of the simulation for the whole network. It is simply calculated by reducing the number of packets received by the application servers from the number of packets generated by the application clients. The network throughput is calculated at application layer as the CBR throughput, which takes advantage of the default calculation method provided by the CBR application in QualNet. The average endto-end packet delay is also calculated using the default method in QualNet. All the values for each traffic flow are averaged at the end as the final packet delay for the network. In addition, we also use Coding Gain to demonstrate how many transmissions can be saved in the network by using opportunistic network coding. Coding Gain is the ratio of number of data packets to the number of transmissions. In this thesis, the data packets are considered as the least priority IP packets dequeued by any nodes in the network since the traffic is carried with UDP protocol. In this situation, even one original data packet is relayed 5 times among a multi-hop network, it should be considered as 5 different data packets because they have different IP headers. On the other hand, the transmission is considered as MAC layer data frame excluding other CTS/RTS transmissions. This statistical information is easily obtained through QualNet simulation platform. Based on 48 these assumptions, the Coding Gain should be 1 in the traditional situation where normally one IP data packet requires one MAC layer data frame, while with network coding, Coding Gain would be greater than 1 because one MAC layer data frame may contain several encoded IP data packets. Coding Gain is a useful method to understand how network coding could improve network throughput. 5.2 Simulation Results Figure 13 shows the simulation results for the Alice-and-Bob topology. As can be seen from Figure 13(a), for the situation without network coding, the throughput of the network increases with actual load and then saturates at 600 Kbps which indicates the capacity of this simple network. For the situation with network coding, the results of three different waiting scheme queue thresholds N, i.e. 0, 1 and 37, are shown. When N is 0 and 1, the corresponding lines coincide with that without network coding. However, the network capacity is improved by 17% to nearly 700 Kbps when the queue threshold N is set at 37. The improvement can be easily explained by Figure 13(b), which shows there is no coding gain when N is 0. In this situation, the waiting scheme does not take action, and thus the relay forwards the input packet immediately when it is received, resulting in no coding opportunity. In addition, for such a simple topology, even in the heavy load situation there would be no more than one packet queuing at the output queue of the relay unless applying the waiting scheme. When N is 1, the coding opportunity is low at the light load situation and almost zero at heavy load. As mentioned earlier, there would be only two packets queuing at the output queue of the relay due to the waiting scheme, resulting in the limited potential coding opportunity. However, in such a wireless channel, Alice and Bob have to contend for the transmission opportunity, which leads to 49 asymmetric traffic flows between Alice and Bob. The asymmetric phenomenon is more common at heavy load situation, which explains the trends shown on the figure. However, the coding opportunity is significantly improved when N increases to 37. It is because there are more packets queuing at the relay’s output queue due to the large queue threshold. The increase of coding opportunity reduces the total transmissions required, resulting in an improvement of network capacity. Correspondingly, the total number of packets dropped is also reduced when N is 37, as shown in Figure 13(d). As expected, the waiting scheme increases the average end-to-end delay under light load situations while the increase in network capacity compensates for the packet delay at heavy load situation, as seen in Figure 13(c). The average end-to-end packet delay is even reduced when the improvement of network capacity is much greater. (a) Comparison of the network throughput 50 Coding gain Delay (s) (b) Comparison of the Coding Gain with different queue threshold N of the waiting scheme (c) Comparison of the average end-to-end packet delay 51 (d) Comparison of the number of packets dropped Figure 13: Simulation results for the Alice-and-Bob scenario showing comparisons of the network throughput, total number of packets dropped, average end-to-end packet delay and coding gain between the situations without network coding and with network coding with different queue thresholds of waiting scheme Figure 14 shows the simulation results for the Cross topology. As can be seen, the network capacity increases by nearly 3 times from 200 Kbps to 800 Kbps because of our opportunistic network coding scheme irrespective of the value of the queue threshold of the wait scheme. Correspondingly, the total number of packets dropped in this network is also significantly reduced. Likewise, as shown in Figure 14(c) the overall average end-toend packet delay is also reduced by the opportunistic network coding scheme though the delay slightly increases due to the waiting scheme at light load. However, in X topology the network coding scheme actually results in poorer network performance compared to the scheme without network coding, as shown in Figure 15(a). Figure 15(b) shows that there is no coding opportunity at heavy load 52 situation in X topology, resulting in no network throughput improvement. The key difference between X and Cross scenarios is that Hello message is the dominating method to carry the reception report in X scenario. Thus the improper value of the Hello message Coding (a) Comparison of the network throughput (b) Comparison of the coding gain 53 Delay (s) (c) Comparison of end-to-end packet delay (d) Comparison of the packets dropped Figure 14: Simulation results for the Cross scenario showing the comparisons of network throughput, total number of packets dropped and average end-to-end packet delay between the situations without network coding and with network coding with different queue thresholds of waiting scheme interval T0 could causes this poor performance. To verify this, we did another simulation on X topology by adjusting the value of T0 from 5 milliseconds to 40 milliseconds to study the impact of Hello message interval on network coding performance. The results are shown in Figure 16. 54 Coding gain (a) Comparison of the network throughput Delay (s) (b) Comparison of the coding gain (c) Comparison of the number of packet delay 55 (d) Comparison of the number of packets dropped Figure 15: Simulation results for the X scenario showing the comparisons of network throughput, total number of packets dropped, average end-to-end packet delay and coding gain between the situations without network coding and with network coding with different queue thresholds of waiting scheme Figure 16(a) shows the coding opportunity in X topology decreases as the Hello message interval T0 increases. If T0 is large, the information exchanged via Hello messages would be out of date, which is useless for opportunistic network coding decision, resulting in less coding opportunity. On the other hand, if T0 is small, nodes could exchange information in a timely manner via Hello messages thus contributing to making useful coding decisions. However, in this case the periodic transmission of Hello messages may consume considerable bandwidth and may even flood the network if the interval is too small. Figure 16(d) show there is an optimal value for T0 to obtain the best performance. The optimal value is 13 ms and it closes to the average time the source sends a packet, which implies replying a node by Hello message immediately after receiving a packet from it is the best arrangement. The comparison of the throughput is shown in Figure 16(b). As can be seen, network coding with optimal T0 improves the network capacity by nearly 25%. The improvement is significantly less than that of Cross scenario because the periodic transmission of Hello messages consume significant bandwidth of 56 the wireless channel. In addition, the average end-to-end packet delay could be also improved by the proper setting of the Hello message interval as shown in Figure 16(c). (a) In X topology where some nodes exchange information via Hello message, the coding opportunity decreases at the increase of Hello message interval. (b) Comparison of the network throughput of X topology between without network coding and with network coding with different Hello message interval T0. The queue threshold N of the waiting scheme is 16. 57 (c) The end-to-end packet delay changes according to the value of Hello message interval. It shows there is an optimal value for Hello message interval in this simple X topology. (d) The network throughput changes according to the value of Hello message interval. It shows there is an optimal value for Hello message interval in this simple X topology. Figure 16: The impact of Hello message interval on the performance of network coding for simple X topology. The above results demonstrate that this opportunistic network coding scheme could potentially improve the network capacity. The study of the waiting scheme and the impact of Hello message interval provide good suggestions on how we should make the algorithm more flexible so that it could be made more suitable for a dynamic neighborhood environment such as in wireless mobile ad-hoc networks. This is especially so for setting the value of the waiting scheme and Hello message interval. Having experimented with these simple topologies, we continue to run the algorithm on a multi-hop scenario. 58 Figure 17 shows the simulation results on a 20-node multi-hop scenario. The figure implies that the performance of our scheme is much poorer when compared to the situation without network coding. Through comparing with the simple Alice-and-Bob scenario, we can say it is caused by the interference of periodic Hello messages. Nodes 1, 3 and 5 communicate with each other, while all other nodes in the network are proactively sending Hello messages among themselves, which in this case seriously interferes the communication channel between nodes 1 and 5, resulting in the poor performance observed. In a large scale network the periodic Hello messages consume large amounts of bandwidth and cause serious channel congestion problem. In order to address this problem, we conclude that the opportunistic network coding scheme should be adapted in a more intelligent way for it to work well. Figure 17: Comparison of the network throughput between without network coding and with network coding with the queue threshold N as 16 on the 20node multi-hop network scenario 5.3 Discussion Simulation results for the simple topologies conclude that the opportunistic network coding scheme is a practical technique, and can potentially improve the network 59 capacity by 3 to 4 times for UDP traffics. In addition, the waiting scheme is a useful feature to this algorithm; however, it depends largely on a number of neighborhood situations. The waiting scheme would be helpful for those networks without bottlenecks, for example the simple Alice-and-Bob scenario. It may help some packets queuing at the relay’s output queue resulting in the improvement in coding opportunity. However, for those networks with bottlenecks such as Cross scenario, the waiting scheme is somewhat redundant, since the packets automatically queue at the bottleneck in heavy load situation. Therefore, it is better if the algorithm is intelligent enough to dynamically apply this waiting scheme based on the actual neighborhood conditions in multi-hop networks. Moreover, the use of Hello message is also affected by the actual neighbor situation and traffic pattern. In the X scenario, the source nodes share information via data packets while the sink nodes use Hello messages. However, in the Cross topology all nodes use data packets to share information. The difference between the results of X and Cross topologies implies that the most ideal situation is for all nodes to exchange information via data packets instead of Hello messages. This is because the periodic Hello messages increase overhead and introduce interference to nodes in the network. So a more practical algorithm should detect the neighborhood situation and local traffic pattern and then make smart coding decisions based on this local one-hop information. What is more, in situations where some nodes have to exchange information by Hello messages, the interval of these Hello messages impacts the performance of this algorithm significantly. Through our simulation, we find there is an optimal value for the Hello message interval which minimizes interference, i.e. it is best for one node to share information immediately after it receives new packets. In the multi-hop scenario with very complex local traffic 60 pattern, it is better for the nodes to adjust the value of this Hello message interval intelligently in order to get the best performance. Thus a more flexible and “intelligent” opportunistic network coding scheme should be proposed to address these issues. The intelligence of the algorithm should contain these competences discussed below. For example, through intelligently listening to the packets from neighbors, the node in the network should make a decision whether to turn on or off the Hello message, because in the multi-hop scenario some nodes do not contribute to the communication at all. In addition, if the node turns on Hello message, it should know how to adjust the value of Hello message interval and how to set the value of the queue threshold for the waiting scheme based on the knowledge of its neighbors and the local traffic pattern. In this way every node would minimize the interference of its Hello messages without sacrificing on coding opportunity. In next chapter, we will present an intelligent version of this opportunistic network coding scheme incorporating some of these suggestions. Further simulations and evaluations are also carried out. 61 Chapter 6 Intelligent Opportunistic Network Coding As discussed in the previous chapter, there are some issues in current opportunistic network coding scheme for wireless environment, such as the ineffective Hello messages. Sending of Hello message is an important technique for the opportunistic mechanisms adopted in this network coding scheme, which makes the nodes in the network able to share information with neighbors thus contributing to overall coding opportunities. However, in the current algorithm the Hello message is scheduled in a proactive style, which means all nodes in the network broadcast Hello messages actively to share information without considering whether the information benefits the coding decision in the relay node or not. This proactive Hello message technique consumes considerable wireless channel bandwidth, thus resulting in poor network performance as observed. In addition, the current network coding scheme does not provide a way to dynamically adjust the value of Hello message interval. As implied in previous simulation results, there is an optimal Hello message interval to achieve best network performance in simple X topology, and this optimal value varies in different network topologies with different traffic patterns. To address these findings, we present our intelligent version of this opportunistic network coding scheme in this chapter. The coding opportunity is detected automatically while the Hello message is scheduled according to an On-demand style with the interval set accordingly. 6.1 Intelligent Hello Message First, we discuss some challenges to make this opportunistic network coding scheme intelligent to address issues mentioned above. In a large scale wireless network, 62 not all nodes are active, which means that some nodes snooping at the network do not generate traffic, forward traffic or receive traffic. These nodes do not do network coding. Even for the active nodes, not all of them have the coding opportunity. As we know this opportunistic network coding scheme only combines the packets from different traffic flows. So only the relay nodes with at least two different traffic flows crossing them have the coding opportunities. However, the issue is how one node in the network is able to discover itself as a relay node with different traffic flows crossing? Likewise, not all information of a node in the network is able to benefit the overall network coding. For example, those inactive nodes which just snoop at the network should not broadcast Hello messages to consume the channel bandwidth since they do not join the communication. In another words, only the neighbors of the relay node with coding opportunities could possibly contribute to the coding decision. However, they should not broadcast Hello messages all the time since the relay node does not always have the coding opportunity. So this is the issue when the neighbors should turn on Hello messages and share their information with the relay node to contribute the coding decision. Furthermore, when the Hello message is turned on, what value should be chosen as the interval? In addition, the coding opportunity does not exist indefinitely, since one network traffic flow would end at any moment while another begins somewhere else. This is another challenge for nodes to decide when to turn off Hello messages. As can be seen, our proposed scheme should be made more intelligent to overcome these challenges and thus improve the overall throughput. The control flow of our intelligent scheme is shown in next section and solutions to address these challenges are presented. 63 6.2 Control Flow In this section, the overall control flow of our intelligent opportunistic network coding scheme is shown. The overall control flow consists of two parts, Host Side and Neighbor Side. The Host Side includes the decision procedures at the relay node which detects the coding opportunity and broadcasts control messages to neighbors. While the Neighbor Side indicates the response procedures at the corresponding neighbors which receive the control messages from the Host Side. Through the communications between the host and neighbors, the neighbors schedule the Hello message in an On-demand style which is more efficient and causes less overhead than the basic approach. 6.2.1 Host Side Figure 18 shows the control flowchart of the Host Side. Initially, all nodes work on promiscuous mode and snoop at the network but with the Hello messages turned off, which is different from the basic algorithm previously. In that previous algorithm, all nodes eavesdropping at the network are eager to share their own states and do not consider whether their information is valuable or not. Here we make all nodes “keep quiet” until some of them are told to share information. This is how our scheme works, whenever a node is ready to send packets, it checks the information of all packets in its output queue, by which it may decide whether it has coding opportunity or not. The information of the packets contains the previous hop, next hop and the destination of this packet. This information can be provided by integrating this opportunistic network coding scheme with some routing protocols. In addition, recall that each node also exploits a waiting scheme with a queue threshold N when it is ready to send packets, which means there are at least N packets in the node’s output queue. By collecting the information array of these N 64 packets, the node is able to detect the traffic pattern crossing itself. The detection procedure is as follows. All packets from the same previous hop and to the same next hop are classified into one traffic flow, since this opportunistic network coding scheme does one-hop network coding. This means that the coding packet will be decoded immediately at the next hop and the next hop will check the coding opportunity for this decoded packet again. Thus, after the node gets the information array of all N packets in the output queue, it may know how many different traffic flows are crossing it. It is easy to determine that a coding opportunity exists when at least two different traffic flows are crossing each other, this is because this network coding scheme only does inter-flow coding. However, the real coding opportunity still depends on the exact topology this node belongs, as it will have a bearing on the overall performance as shown in the simulation results presented in chapter 5. In order to learn its local topology, the node should have information up to two hopes away, which can be obtained by a route discovery protocol which we will not elaborate here. Instead, we just simply assume all nodes are able to discover the two-hop away information in current configuration. The explicit route discovery protocol design and integration with this network coding scheme is beyond the scope of this work. Once the node finds out its local topology, a decision metric to determine its coding decision is available. For example, take any two traffic flows, if the next hop of one traffic flow can hear the previous hop of another traffic flow and similarly for the other pair, then packets from these two flows can be combined together since the next hops of these packets are able to decode the combined packets. This is easily illustrated in the X topology shown in Figure 11(a) in chapter 5. There are two traffic flows crossing the relay node, from node 1 to node 5 and from node 2 to node 4, which can be written as 1->5 and 2->4. As long as 65 {1, 4} and {2, 5} are able to hear each other, we deem that the relay node has coding opportunity. This decision metric can also be applied in Alice-and-Bob, Cross or other scenarios. Figure 18: Host Side Once the node has coding opportunity, it continues to check whether it needs the help of its neighbors since the node still needs the reception report from neighbors to do network coding. The procedure described here shows how the node decides whether it needs help or not. First, if the node has no records for the intended neighbor, the node 66 must need the reception report from the neighbor. On the other hand, if the intended neighbor has previously sent something to this node, this node must have recorded the information in its local buffer. So this node checks the time stamp of the neighbor’s record in local buffer. If the record is relatively new then this node does not need the help from the neighbor because the neighbor has sent the latest information to it. Otherwise, this node would need the new information from the neighbor. If the time stamp of the record is within the last T2 seconds called New Record Duration, the record is regarded as new; otherwise, it is regarded as old. If the node needs the help from neighbors, it just schedules a control message broadcast to the neighbors informing them to turn on Hello message with a specific packet interval T0. This T0 is the average input packets interval of the traffic flow which can be measured when the node is checking the packets’ information in the output queue. Once the above coding decision is taken the node goes through the opportunistic coding procedures as previous algorithm. On the other hand, if this node has no coding opportunity, it will simply skip the opportunity coding procedures. Before that, it has to check whether it has informed its neighbors to turn on Hello message, if it did, it should schedule another control packet to the neighbors asking them turning off the Hello message. 6.2.2 Neighbor Side Figure 19 shows the flowchart for the Neighbor Side, where the procedure is much simpler since the Neighbor Side just does what the Host tells them to do. When the neighbors receive control packet from the relay node, they simply check the flag in the control packets to decide whether to turn on or off the Hello message. If they have to turn on, they just schedule the corresponding periodic timer based on the Hello message 67 interval enclosed in the control packet. If they have to turn off, they just simply cancel the Hello message timer maintained in the protocol state structure. In addition, when the neighbors are ready to broadcast Hello messages, they should always check whether they have received new messages during the last T1 second called New Packet Duration. If not, they should stop broadcasting Hello messages; otherwise, they broadcast the Hello message as usual. This scheme prevents the neighbors from broadcasting useless information to the relay node, which may again consume extra wireless bandwidth. Figure 19: Neighbor Side We have presented the intelligent version of the opportunistic network coding scheme in this chapter. In next chapter, the evaluation of this intelligent scheme will be elaborated. 68 Chapter 7 Simulation And Evaluation In previous chapter, we have presented the details of our intelligent version of opportunistic network coding for wireless mobile ad-hoc network. In this chapter, simulation and evaluation of this intelligent scheme are studied and presented. The implementation of this intelligent scheme is relatively simple, and we just add a few extra protocol states in the CopeData structure. An extra coding opportunity detection function is added before the coding procedure function Coding(). In addition, a control packet handler function is implemented to process the control message at the neighbor side. The details of the implementation are found in our source code. With regards to the simulation, most of the simulator parameters and performance metrics remain as the same as those mentioned in chapter 5. We further expand our simulation from simple wireless topologies to large scale multi-hop wireless environment, thus evaluating the performance of our intelligent scheme in large scale networks. In our simulation, the New Packet Duration T1 and New Record Duration T2 are both set as 1 second. 7.1 Simulation It may be better to compare the performance of this intelligent algorithm with the previous basic algorithm on those simple scenarios shown in Chapter 5. However, this intelligent algorithm is based on the previous basic one and deals well with large scale wireless networks, which means actually the same simulation results are expected for both algorithm on those simple network scenarios. Thus we do not present those simulation results on the simple scenarios here due to the space limitation. In fact, those simulation results shown on Figure 20 and 23 can be the very evidence for this assumption. 69 We simulated the intelligent scheme in the same 20-node scenario shown in Figure 12 in chapter 5 to show the expected improvement based on the previous basic algorithm. In this scenario, node 1 and node 5 communicate with each other with the help of node 3, where the route is discovered by AODV. Figure 20 shows the simulation results on this scenario. As can be seen from Figure 20(a), the lines with square markers and diamond markers indicate the network throughput achieved by non-coding method and basic coding algorithm respectively, which have been shown in Figure 17. The basic network coding scheme makes the network performance much worse than the non-coding method on a large scale wireless network. However, the line with x marker indicating the network throughput achieved by our intelligent scheme shows the performance of the network has significantly improved compared to the basic coding scheme. The network throughput achieved by intelligent scheme is also much better than that of non-coding method, which has been improved by nearly 25%. What is more, as shown in Figure 20(c), the total number of packets dropped has also been significantly reduced by our intelligent scheme in this large scale wireless network. However, this improvement in network throughput is achieved with larger average end-to-end packet delay. Figure 20(b) shows the average end-to-end packet delay has increased, especially at the light load situation. The trimodal delay curve is expected because in this simple traffic pattern one packet would have to experience two stages of delay due to the waiting scheme. The first stage of delay happens at the output queue of the source while the second happens at the output queue of the relay node. At each stage the delay can be modeled like this: I# ˠ >ˠ ˤ# = {˚ ∗ ˠ ˠ ≤ˠ 70 I$ ˠ > ˠ ˤ$ = {˚ ∗ ˠ   ˠ ≤ˠ where d1 and d2 are the delay caused in source and relay node. c1 and c2 are the constant delay when the waiting scheme is not executed. N is the threshold of the number of packets for the waiting scheme. T is the waiting duration for the waiting scheme. Here it is set at 40 milliseconds. Ti is the input packet interval at source while Ti’ is at the relay node. It is easy to find that Ti=2Ti’. Thus d2 could be written as: I$ ˠ > 2ˠ ˤ$ = {˚ ∗ 0.5 ∗ ˠ ˠ ≤ 2ˠ Then the total delay d is I# + I$ ˤ = ˤ# + ˤ$ = {I# + ˚ ∗ 0.5 ∗ ˠ ˚ ∗ 1.5 ∗ ˠ ˠ > 2ˠ ˠ < ˠ ≤ 2ˠ ˠ ≤ˠ which is obviously a bimodal curve matching the first half of the one with x marker shown in Figure 20(b). The tail of the curve is caused by the significant packet drop which would lead the longer and longer average end-to-end packet delay. Besides this simulation on the simple traffic pattern, we expand our simulation to more complex multi-hop traffic patterns. As mentioned in the previous chapter, we do not design the specific route discovery protocol to discover two-hop away information required by our intelligent scheme but simply assume that all nodes in the network have two-hop away information. In order to simplify the procedure to set the static route table for each node, we simply exploit a regular multi-hop wireless scenario shown in Figure 21 where all 20 nodes are regularly placed in a grid covering an area of 1200m*1200m. In this scenario, each node can only hear its surrounding neighbors and the static route table for each node is set according to this. 71 (a) Comparison of network throughput (b) Comparison of average end-to-end packet delay (c) Comparison of total packet dropped Figure 20: Simulation results for the 20-node scenario showing the comparisons of network throughput, total number of packets dropped, average end-to-end packet delay between the situations without network coding, with network coding and intelligent network coding 72 Figure 21: 20-node multi-hop scenario, where 20 nodes regularly placed in a grid area covering 1200m*1200m We first repeat the simple X and Cross scenarios from the original 5-node topology to this 20-node environment which is shown in Figure 22. As expected, the simulation results on this larger scale wireless network are similar to that of the simple scenarios. This is proven to be the case by the simulation results shown in Figure 23. Figure 23(a) shows the network throughput is improved by nearly 40% from 500 Kbps to 700Kbps due to the intelligent scheme. Figure 23(b) shows for the Cross topology the network throughput is improved by almost 2.7 folds from 300 Kbps to 800 Kbps. 73 (a) X scenario (b) Cross scenario Figure 22: X and Cross scenarios on regular 20-node wireless environment (a) Comparison of network throughput on X scenario 74 (b) Comparison of network throughput on Cross scenario Figure 23: Network throughput from X and Cross scenarios on regular 20-node wireless environment (a) 4-hop chain scenario (b) Long X scenario Figure 24: Multi-hop scenarios on regular 20-node wireless envrionment Next we further expand our simulation to multi-hop traffic flows. Figure 24(a) shows the 4-hop away chain scenario where the node 6 and node 10 communicate with 75 each other by CBR under the help of the relay nodes between them. This scenario is similar to the simple Alice-and-Bob scenario but it is two-hop further. Thus, the coding procedures in this scenario would be much more complex. The simulation results are shown in Figure 25 where we see that the total numbers of packets dropped on these two different methods are almost the same while the network throughput achieved by our intelligent network coding scheme is a bit better than that achieved by non-coding method. Another example is the Long X scenario shown in Figure 24(b) which is expanded from the simple X scenario. However, the overall coding procedure is quite different from that of the basic scenario. For example, the other nodes except those circled would decide to keep quiet without sending Hello message during the communication period since they know they are not able to contribute to the ongoing communication. In addition, the relay node with two traffic flows crossing should decide not to ask its neighbors to turn on Hello message while the other two relay nodes, 8 and 14, should detect that they have no coding opportunity and thus they do not to ask the two destinations to turn on Hello messages either. All these features were not available in the previous opportunistic network coding algorithm. The corresponding simulation results are shown in Figure 26 where we observe that the maximum network throughput is improved by nearly 33% from 300Kbps to 400Kbps compared to the non-coding method. Meanwhile, the total number of packets dropped is also significantly reduced by the intelligent scheme. 76 (a) Comparison of network throughput (b) Comparison of total number of packets dropped Figure 25: Simulation results for the 4-hop chain scenario (a) Comparison of network throughput 77 (b) Comparison of total number of packets dropped Figure 26: Simulation results for the Long X scenario 7.2 Evaluation As can be seen from the simulation results shown above, our intelligent opportunistic network coding scheme manages to solve those issues encountered in the basic algorithm. Our intelligent scheme is able to work well in the large scale multi-hop wireless environment by changing the proactive Hello message technique with on-demand Hello message. Those nodes which cannot contribute positively to the network coding communication keep quiet without causing any overhead and interference to the network. Other nodes which take part in the communication procedure are able to detect coding opportunity and correctly indicate to their neighbors to turn on or off Hello messages. For example, in the 4-hop away chain scenario, our simulation results show that all three intermediate nodes, 7, 8 and 9, have coded several packets. While in the Long X scenario, only node 13 has coded packets and the other two nodes, 8 and 14 just forward packets because of no coding opportunity. Therefore, the two destination nodes just receive the data quietly without broadcasting any Hello message. 78 Though only some regular wireless scenarios are chosen to verify our intelligent opportunistic network coding scheme, the simulation results are sufficient to show that our intelligent scheme is able to improve overall network throughput no matter what the topology is albeit at different degrees. For some specific network topologies and traffic patterns, the network throughput could be improved up to 3 folds. Our intelligent scheme is effective in detecting potential coding opportunity based on the local topology and the current traffic patterns, thus taking corresponding actions to maximize overall network throughput. Generally, the total number of packets dropped is also significantly reduced. In next chapter, we will present our overall conclusions arising from the implementation and evaluation of our opportunistic network coding scheme in wireless mobile ad-hoc networks. Future work will be also outlined to facilitate possible enhancements to our intelligent opportunistic network coding scheme. 79 Chapter 8 Conclusions And Future Work 8.1 Conclusions Network coding has been demonstrated to be a promising technique in current wireless packet switched networks. In order to bring the benefit of network coding to the wireless mobile ad-hoc networks we have designed an intelligent opportunistic network coding scheme and implemented it in QualNet simulator. To achieve this, we first study the behavior of opportunistic network coding scheme and evaluate its performance in wireless mesh networks filled with UDP traffic. Our implementation and subsequent investigations using QualNet simulator has shown that this intelligent opportunistic network coding scheme is practical to be integrated into the current protocol stack and work well with current 802.11 wireless networks. What is more, by solving the overhead and interference caused by inefficient Hello messages, this intelligent network coding scheme can be implemented in a large scale network. Moreover, our simulation results show this intelligent opportunistic network coding is able to improve the overall network throughput of wireless mesh networks with UDP traffic. The total number of packets dropped is also significantly reduced. Though the improvement in performance highly depends on the overall wireless scenario and exact traffic patterns, and the overall network throughput of some specific simple topologies can be improved by nearly 3 fold with significant reduction of packets dropped. Even for long chain scenario the network throughput can be improved up to 10%. 80 8.2 Future Work Our target is to research network coding in wireless mobile ad-hoc networks (MANETs), to improve parameters such as the network throughput gain and robustness. Our work here has achieved some important milestones with respect to this goal, and also shed light on possible future works which will complete the overall objective. First, in order to evaluate the performance of pure network coding we have made our intelligent network coding scheme to be independent of the underlying routing protocol. In future therefore, an explicit routing protocol for wireless mobile ad-hoc network which is able to discover the local topology should be designed and integrated with our intelligent network coding scheme for evaluation. We believe by taking full advantage of efficient routing protocol for wireless mobile ad-hoc network, the intelligent opportunistic network coding scheme proposed should have better performance in mobile environments. In addition, the impact of two parameters during the coding opportunity discovery procedure i.e. New Packet Duration and New Record Duration should be evaluated further, especially in mobile environments. These two parameters may be associated with the degree of mobility of the network nodes. On the other hand, we have evaluated the performance of network coding based on UDP traffic, therefore, the more complex TCP traffic can be exploited next. Compared to UDP protocol, TCP traffic contains more features such as congestion control, acknowledgement and retransmissions, which may result in additional challenging issues to be addressed. Finally, another interesting area for consideration would be information security. In our intelligent opportunistic network coding scheme, we not only exploit simple XOR to combine input packets but also allow all nodes to eavesdrop in the network, which may 81 be a potential security issue. A study on the security issue of network coding for wireless mobile ad-hoc network would be relevant here. 82 REFERENCES [1] R. Ahlswede, N. Cai, S. Li, and R. Yeung, “Network information flow,” IEEE Transactions on Information Theory, vol. 46, no 4, Jul. 2000, pp 1204-1216. [2] S. Katti, H. Rahul, W. Huss, D. Katabi, M. M´edard and J. Crowcroft, “XORs in The Air: Practical Wireless Network Coding,” ACM SIGCOMM, Pisa, Italy, Sep. 2006. [3] Qunfeng Dong, Jianming Wu, Wenjun Hu and Jon Crowcroft, “Practical Network Coding in Wireless Networks”, ACM MOBICOM, Montréal, Québec, Canada, Sep. 2007 [4] Christina Fragouli, Dina Katabi, Athina Markopoulou, Muriel M´edard and Hariharan Rahul, “Wireless Network Coding: Opportunities & Challenges” MILCOM IEEE, Orlando, FL, USA, Oct. 2007, pp. 1-8. [5] S. Katti, D. Katabi, W. Hu, H. S. Rahul, and M. M´edard. “The importance of being opportunistic: Practical network coding for wireless environments”, 43rd Annual Allerton Conference on Communication, Control, and Computing. Monticello, IL., USA, Sept. 2005. [6] S. Li, R. Yeung and N. Cai, “Linear network coding”, IEEE Transactions on Information Theory, vol. 49, no. 2, Feb. 2003, pp. 371–381. [7] R. Koetter and M. Medard, “An algebraic approach to network coding”, IEEE/ACM Transactions on Networking, vol. 11, no. 5, Oct. 2003, pp. 782–795. [8] T. Ho, M. Medard, J. Shi, M. Effros, and D. Karger, “On randomized network coding”, 41rd Annual Allerton Conference on Communication, Control, and Computing. Monticello, IL., USA, Oct. 2003. [9] J-S. Park, M. Gerla, D. S. Lun, Y. Yi and M. Medard, “CodeCast: A Network-CodingBased Ad Hoc Multicast Protocol”, IEEE Wireless Communications, vol. 13, no. 5, Oct. 2006, pp. 76-81. [10] A. Ramamoorthy, J. Shi and R. Wesel, “On the capacity of Network coding for random networks,” IEEE Transactions on Information Theory, vol. 51, no. 8, Aug. 2005, pp. 2878–2885. [11] S. Deb, M. Effros, T. Ho, D. Karger, R. Koetter, D. Lun, M. Medard and N. Ratnakar, “Network coding for wireless applications: A brief tutorial,” International Workshop on Wireless Ad-hoc Networks (IWWAN), London, UK, May 2005. [12] D. Lun, N. Ratnakar, R. Koetter, M. Medard, E. Ahmed and H. Lee, “Achieving minimum-cost multicast: A decentralized approach based on network coding,” IEEE INFOCOM, Miami, FL, USA, Mar. 2005. pp. 1607-1617 83 [13] S. Chachulski, M. Jennings, S. Katti and D. Katabi, “Trading structure for randomness in wireless opportunistic routing,” ACM SIGCOMM Kyoto Japan, Aug. 2007. [14] S. Katti, D. Katabi, H. Balakrishnan and M. Medard, “Symbol-level network coding for wireless mesh networks,” ACM SIGCOMM. Seattle WA, USA, Aug. 2008. [15] X. Zhang and B. Li, “Dice: A game theoretic frame work for wireless multipath network coding,” 9th ACM International Symposium on Mobile Ad-hoc Networking and Computing, Hong Kong, China, May 2008, pp. 293-302. [16] Christina Fragouli, Jean-Yves Le Boudec and Jorg Widmer, “Network Coding: An Instant Primer,” ACM SIGCOMM Computer Communication Review, vol. 36, no. 1, Jan. 2006, pp. 63-68. [17] Internet packet size distributions: Some observations. http://www.isi.edu/~johnh/PAPERS/Sinha07a.html http://netweb.usc.edu/~rsinha/pkt-sizes/. [18] D. S. DeCouto, D. Aguayo, J. Bicket and R. Morris, “A High Throughput Path Metric for Multi-Hop Wireless Routing,” 9th Annual International Conference on Mobile Computing and Networking, San Diego, CA, USA, Sep. 2003, pp. 134-146. [19] C. E. Perkins, E. M. Royer, “Ad-hoc On-Demand Distance Vector Routing” 2nd IEEE Workshop on Mobile Computing Systems and Applications, New Orleans, LA, USA, Feb. 1999. [20] T. Clausen, P. Jacquet, A. Laouiti, P. Muhlethaler, A. Qayyum and L. Viennot. “Optimized link state routing protocol for ad hoc network” IEEE International Multi Topic Conference, Pakistan, Dec. 2001, pp. 62-68. [21] Ajoy Kar. “Performance study of practical wireless network coding in 802.11 based wireless mesh networks” Master Thesis, National University of Singapore. 2009 84 [...]... coding scheme in QualNet As mentioned in chapter 2, this opportunistic network coding scheme works between the MAC and IP layers by inserting an extra coding header between MAC and IP headers However, in order to be consistent with the real network architecture, we do not create an extra independent communication layer between Link and Network layer Instead, we implement the opportunistic network coding. .. receive the message in the layer short instanceId; short eventType; // Message's Event type …… MessageInfoHeader infoArray[MAX_INFO_FIELDS]; int packetSize; char *packet; NodeAddress originatingNodeId; int sequenceNumber; int originatingProtocol; int numberOfHeaders; int headerProtocols[MAX_HEADERS]; int headerSizes[MAX_HEADERS]; Figure 6: Message structure in QualNet Some of the fields of message data... implementation 3.1 Packet header Figure 4 shows the modified variable-length coding header for the opportunistic network coding scheme, which is inserted into each packet If the routing protocol has its own header, our coding header sits in between the routing and MAC layer headers Otherwise, it sits between the IP and MAC headers Only the shaded fields in Figure 4, is required in every coding header (called the... wireless environment In next chapter, we will show the specific architecture of opportunistic network coding in our implementation including the specific packet header structure and the whole control flow of the algorithm 21 Chapter 3 Opportunistic Network Coding Architecture In this chapter, we introduce the architecture of the opportunistic network coding scheme which we have implemented in QualNet simulator... the intended packet and then stores a copy of the decoded native packet in its packet pool As all packets are sent using 802. 11 unicast, the MAC layer is able to detect collisions and back-off properly Pseudo broadcast is therefore more reliable than simple broadcast and inherits all the advantages of broadcast In this chapter, we have presented the overview of the opportunistic network coding for wireless. .. specific destination at a time However, unicast does not help in opportunistic listening and consequently not in opportunistic coding To address this problem pseudo broadcast is introduced in COPE Pseudo broadcast is actually unicast and therefore benefit from the reliability and the back-off mechanism The link layer destination field of the encoded packet is set to the MAC address of one of the intended... details of the architecture are based on the overview of the opportunistic characteristics, coding and decoding algorithm introduced in the previous chapter Here the packet header structure will be shown and the functionality of each field in the header will be explained Next the overall control flowchart will be presented to illustrate the structure of the opportunistic network coding algorithm in our implementation. .. packet reordering as a congestion signal 2.3 Packet Coding Algorithm In this section, we introduce the details of packet coding algorithm based on the opportunistic coding idea mentioned above Some practical issues in the implementation and the specific solutions are also proposed First, the coding scheme in the original COPE does not introduce additional delay The node always de-queues the head of its output... be part of an intelligent scheduling scheme, which will be described at Chapter 6 11 In summary, through the Opportunistic Listening technique, the nodes learn and share their states with each other, thus contributing to the nodes’ network coding decision 2.2 Opportunistic Coding Opportunistic coding allows nodes to combine multiple packets into a single packet based on the assumption that all intended... congestion situation in static wireless mesh networks, we do not implement this technique in our design for several reasons First, we would like to design an opportunistic network coding scheme independent of routing protocol, thus making the algorithm very flexible at difference scenarios by cooperating with the suitable routing protocol instead of integrating with ETX based routing algorithm Furthermore, ... message on the performance of opportunistic network coding scheme We then use the findings for further study of network coding scheme on large scale wireless ad- hoc networks We find out there is an... mobile ad- hoc networks In this thesis we propose to first study and investigate the performance behavior and key control parameters of opportunistic network coding in large scale static wireless. .. illustration of network coding, showing how network coding saves bandwidth consumption It shows Alice and Bob to exchange a pair of packets using transmissions instead of Recently the research focus on network