Joint Approach for IP Overhead Reduction in Wireless Ad Hoc Networks

Transcription

1 Joint Approach for IP Overhead Reduction in Wireless Ad Hoc Networks Tatiana K. Madsen, Frank H.P. Fitzek, Gian Paolo Perrucci and Ramjee Prasad Department of Communications Technology, Aalborg University Neils Jernes Vej, 90 Aalborg Øst, Denmark, [tatiana ff gpp Abstract Wireless ad hoc networks offer attractive features, including flexibility, self-organization and dynamic network creation. Due to the limited bandwidth- and energy-resources in ad hoc networks, minimization of control- and processing overhead is desirable. In this paper we focus on transport overhead reduction by means of header compression techniques. The potential of header compression in ad hoc networks is not fully exploited yet and is mostly based on tag switching approaches or full decoding of the routing information at each node. While the former approach can only be applied to dedicated technologies, the later comes with the price of large processing and energy overhead. In this paper we advocate a new joint approach where compression is closely coupled with ad hoc routing. To keep the processing overhead low we suggest end-to-end compression. Packet routing can be implemented by using layer III-to-layer II routing mapping. This approach facilitates fast forwarding, since compressed IP headers are not examined at every node. Simulation results indicates the transport overhead reduction of about 70 %. I. INTRODUCTION As wireless delivery of voice and data continues to evolve, ad hoc capabilities are expected to become more and more demanded and the future technological solutions should support this development. To cope with dynamic nature of mobile ad hoc networks, the network functionalities should be maintained in a distributed manner. This coupled with bandwidth and energy limitations dictates design of special networking protocols that possess specific features and optimize the use of scarce resources. Scanning a vast amount of ad hoc research literature, one can see that the main optimization criteria are to minimize control overhead, processing overhead and energy consumption. E.g. an efficient routing protocol will keep the overhead and memory requirements low, while being able to establish and maintain a path between a source and a destination in case of topology changes due to node movement. The amount of overhead produced by a protocol is directly related to its scalability. Networking overhead can be divided into signalling overhead and transport overhead. For the transportation of data over an IP based network, the related overhead in terms of additional header information has to be taken into account. We refer to this kind of overhead as transport overhead. For example, delivery of a real-time multimedia or voice over IP with RTP/UDP/IP suit results in 0 bytes IP header for IPv and 60 bytes of overhead for IPv6. Header compression provides us with a tool to reduce this overhead. Header compression schemes have been originally designed for individual links. The potential of header compression in ad hoc networks is not fully exploited yet; only some efforts has been done in this direction. In [] a Routing-Assisted Header Compression (RAHC) technique is proposed. A partial flooding of context information is adapted for RAHC: all nodes that lie in between the source and the destination are updated with the context information depending on the current topology of the network and the path set up by the routing algorithm. Somewhat similar idea is used in [] for SEEHOC (scalable and robust end-to-end header compression techniques for ad hoc networks): a context state update is requested from the upstream node and, when the update is received, it is forwarded to the downstream node. SEEHOC offers a number of solutions for maintaining the context state and detecting bit errors, e.g. for RTP flows incremental encoding is not used, then the loss of a compressed header will not invalidate the context state. A novel idea of maintaining the context states of multiple IP flows in a cooperative manner is suggested in [7]. Cooperative Header Compression () is developed for multi-channel communication systems. This approach requires static routing tables for data forwarding and, therefore, it is focused on a class of networks termed mesh networks. Unlike ad hoc networks, mesh network serve as access networks relaying traffic to and from the wireless Internet in a multi-hop fashion and they consist of immobile nodes. It has been realized that in wireless networks, and especially in ad hoc networks, synergies between the design of different protocol layers must be exploited to improve the system performance []. The advantages of the cross-layer design have been repeatedly demonstrated. Typically, joint optimization of the routing, MAC and link layer is performed for throughput maximization (see e.g. [5]). In this paper our focus is on header compression and, following cross-layer paradigm, we ague for the joint optimization of the compression and routing in order to achieve high compression gain and bandwidth efficiency. We demonstrate that in dynamic multi-hop environment the design of header compression cannot be considered separately from the underlying routing protocol. The presented approach is adopted to specifics of mobile ad hoc networks and it incorporates a number of unique features. The main characteristics of the proposed header compression scheme are: i) no compression/ decompression is performed at the intermediate nodes; ii) to maintain the context state no feedback from the

2 destination is used. To support this compression approach we consider layer III-to-layer II routing mapping. We propose a simple technique that tags each stream with a flow identifier and uses cache information at the intermediate node to find the next hop for the routing. This method, known as label switching, is used for MPLS and ATM virtual circuits. We show that a context ID (CID) can be used as the flow identifier provided different CIDs are assigned to different compressed packet streams. The remainder of the paper is as follows. Section II provides the in-depth description of the proposed approach. Section III discusses the problem of the unique CID assignment in ad hoc networks and presents CID deny messaging method. Performance evaluation is presented in Section IV. Section V gives some concluding remarks. II. PROPOSED APPROACH Traditional header compression approaches, designed for a single hop communication, should be reconsidered for the case of multi hop networks. Since an ad hoc network can consists of heterogeneous nodes, the intermediate nodes along a path between a source and a destination can greatly vary with respect to their processing capabilities and, thus, their offered services. Some of the nodes might not be able to perform header compression. One solution is not to apply header compression over links where such nodes are either a sender or a receiver. Alternatively, header compression/ decompression should be performed only at the end nodes, thus the intermediate nodes will forward packets with the compressed headers. This solution will keep the processing overhead low. It can be easily implemented, unless IP header information is used for routing the packets in the ad hoc network. For example, if MAC addresses are used for routing, there is no need for an intermediate node to have information contained in the full IP header. In case of IP-based routing, we propose to enhance a routing protocol by applying label switching approach. The concept of routing flows using perhop state is a part of Multi-Protocol Label Switching (MPLS) and ATM. In the wireless ad hoc settings this approach is considered for implicit source routing [6]. Designing a compression algorithm, one should be careful about context synchronization between a compressor and decompressor. If an IP packet is lost, it can potentially lead to the loss of the context and the context update is required. Considering end-to-end compression, the context update request has to transverse multiple hops to reach the compressor and, thus, it might not be received timely. In case when only unidirectional links exist between the nodes, a new path should be discovered from the destination to the source node, thus signalling and complexity of the scheme is increased. For these reasons, we have chosen to focus on the header compression schemes that do not rely on the feedback information. As examples of such schemes we take delta coding (see e.g. []) header compression and Cooperative Header Compression (). The detailed description of can be found in [7], [8]. Delta coding is a basic technique in header compression. Correlation in the variation in the field information of the packet flow allows compression of the header to the relatively small sizes, e.g. bytes. Using delta coding, only differential information referring to the preceding header is included in the packet header. In order to establish the context at the decompressor side, first an uncompressed header is sent; it is followed by a row of compressed headers. For error-free channels this is a very efficient way of header compression. However, it is very sensitive to packet losses: if one packet is lost, the context at the decompressor is not updated and all the subsequent packets, even if received correctly, can not be decompressed. In such situations a context update is required. The update can be sent either by a request from the decompressor side or, in case when a feedback is not available or is too costly to provide, periodically. In this paper we consider the later case. The frequency of the updates depends on the channel error rate and the propagation environment. Frequent updates (short frame sizes) keep the system robust, but compression gain is decreased. Long frame sizes increase the probability of loss of the context. There is a trade-off between the robustness of the scheme and its compression gain. Jointly addressing header compression and routing, we consider Dynamic Source Routing (DSR) [9]. Our choice of DSR protocol is due to its simplicity and low signalling overhead. The main disadvantage of DSR is the problem with its scalability when the number of hops between a source and a destination grows. In this situation the amount of overhead increases. But this is due to increase in transport overhead (!) since the DSR header size includes IP addresses of all intermediate nodes along the path and can become very large. Our proposal reduces the header size to or 6 bytes and the compressed header has a constant size regardless of the number of hops between the source and the destination. This makes DSR an attractive choice for ad hoc routing. One should note, though, that our proposal can work with any ad hoc routing protocol provided that it is integrated with label switching. Next section addresses the integration of label switching approach with mobile ad hoc networks routing protocols and assignment of unique labels in a distributed way. III. UNIQUE ASSIGNMENT OF FLOW IDENTIFIER As it has been mentioned before, if a routing protocol, like e.g. DSR, relies on IP header information to route packets in the network, we propose to introduce a flow identifier and integrate it with the existing routing protocols. The idea behind introducing a flow identifier is to use it as the identification of the route to be followed by a particular data stream from a source to a destination. Intermediate nodes maintain Flow Table indicating the next hop the packets with this flow ID should be forwarded. By applying this method, the compressed headers should include a flow identifier as one of their fields and exactly this information will be used for routing. The main problem with label switching in ad hoc networks is how to assign flow IDs in an unique manner. Since there

3 can be several source destination pairs and each source node is choosing a flow ID independently of the others, there can occur situations when the same ID is chosen for different streams. In [6] a tuple source address, destination address, flow identifier is suggested to use as an unique identification of a stream. Using this approach the efficient compression of the packet header fields can not be achieved since IP addresses of the source and destination can not be omitted. In order to achieve high compression gain, it is desirable not to include static information into the compressed headers. We propose to use a context ID as the flow identifier since it does not require addition of any extra information to the compressed packet header. Typically, the first field of the compressed header contains the context ID (CID) that is used to identify which connection the packet is associated with. In static IP networks usually no problems occur with unique CID assignment. But in multi-access and multi-hop networks several compressors and decompressors share the same CID space. CID can no longer be selected independently as collision may occur. One can solve this problem by keeping separate CID spaces for each source-destination pair. Unfortunately, this method requires a large CID space. In this paper we present a solution based on CID deny messaging. Another possible approach is CID swapping presented in []. MAC address MAC address IP flow IP flow MAC address MAC address IP flow MAC address Fig.. a) b) c) CID collision at the intermediate node. 5 MAC address MAC address 5 MAC address MAC address MAC address In case when there are two intersecting IP flows with the same CID, as it is illustrates on Figure a, but the two upstream and the two downstream nodes are different, there is no need to change CID assignment of the flows. Indeed, if the intermediate node (node ) maintains in its Flow Table not only CID, but also MAC addresses of the upstream and downstream nodes, then with the help of link layer information routing can be performed. Thus, CID re-assignment can be avoided in this situation. One can note that the situation, presented on Figure b, can not occur at all. In case node is the source, it will never assign the same CIDs to different IP flows. Otherwise, the CID conflict will be detected already at the intermediate node and the conflict will be resolved before the flow with already used CID will reach node. Only in the situation depicted on Figure c the conflict occurs and it should be resolved. The CID deny message will be send by node to the source requesting to choose another CID. If all the CIDs at a particular node are being used, then the node will reject the establishment of new context until a CID becomes available or after a timeout it will repeat an attempt. Until then, the packets will be sent with an uncompressed headers. To summarize, we would like to underline that the proposed label-switching can be efficiently combined with header compression. Indeed, from the point of view of upper layers, label-switching applied in ad hoc networks masks multi-hop connections as a single-hop link. A. General evaluation IV. PERFORMANCE EVALUATION For a general evaluation, let us assume that we can model each link or hop using uncorrelated bit errors. This is a useful approximation, especially if heavy interleaving is applied. Since the link quality can be different for different pairs of nodes, we denote the probability of a bit sent and not received successfully over a link between nodes i and j as BER ij. Then, the probability to receive an uncompressed header of U bytes is P U ij = ( BER ij) 8 U. Analogous formula is valid for the compressed header. Here we assume that error handling of payload erroneous bits is done by the concealment mechanisms at the application layer. Thus, the packet delivery over multiple hops is increased. This assumption is introduced, since we are interested in the impact of erroneous compressed headers on the performance of the header compression schemes. Note that this is a feasible assumption about the behavior of the networking layer if UDP lite is used. Consider a path L consisting of L hops. Let the random variable γ denote the number of packets received and decompressed correctly in the frame with the length N. For the delta coding approach the expectation of r.v. γ can be found as E[γ] = ( Pij U ) i,j L ( P ij U )N i,j L ( P ij U ) () i,j L To derive formula, we take into account that a packet should transverse multiple hops along the path L to reach the destination. If one packet is lost, then all subsequent packets are lost as well.

4 To find the expectation for approach, one should note that the context loss can happen only if packets on all cooperative channels are lost in one time instant. To simplify formulas, we introduce the following notations: P U(C) = ( P U(C) ij ) () i,j L where P U (P C ) is the probability to receive an uncompressed (compressed) header correctly at the destination. Then for E[γ] = P U + P C ( ( P U ) J ) ( ( ( P C ) J ) N ) ( P C ) J () where J is a number of cooperative channels. Now the important performance metrics, such as packet delivery and decompression ratio (PDDR) and bandwidth efficiency (BE), can be found using the expectation of γ. PDDR is defined as the number of correctly decompressed packets over the total number of packets sent, thus, P DDR = E[γ]/N. Bandwidth efficiency is defined as a ratio of correctly received useful information to the total amount of information sent. PDDR indicates how robust towards the errors the compression scheme is. Bandwidth efficiency reflects both the compression gain and the error rate. Figure illustrates bandwidth efficiency of the delta coding and schemes for different values of bit error rate. Here we assume that all links have the same quality. The point N = corresponds to the case when all packets are sent with uncompressed headers. From Figure we can see a rapid increase in BE when the frame length becomes larger. As the number of packets in a frame continue to increase, BE decreases. One can clearly see that the optimal size of the frame for e.g. delta coding when BER is 0 is around packets. In all cases, the performance of is better compared to delta coding. This illustrates the advantage of cooperative behavior of the routes. Bandwidth efficiency Fig.. Delta, BER=0, BER=0 Delta, BER=0, BER=0 Delta, BER=0 5, BER= Frame length Bandwidth efficiency as a function of frame length. In this evaluation we have assumed that all packet errors are due to the channel conditions but not due to the mobility of nodes. To take into account node mobility, along with other factors, simulation studies are performed. Their results are presented in the next section. B. Simulation results The implementation is done using ns- simulator [0]. This simulator is easily available public domain tool and is considered a de facto standard for implementation and verification of ad hoc networks. Since packets generated by ns- agents do not carry the full headers, but only source and destination addresses, generation time-stamps, type and size of the packets, we have chosen to perform IP header compression/ decompression outside ns- (see Figure ). Fig.. Implementation. Simulation is carried out with 50 nodes randomly distributed over an area of m. Nodes are moving according to the random waypoint model. Constant bit rate flows over UDP are set up with a nominal bit rate 0. bps. Payload of 0 bytes is considered. Free space model is chosen as a propagation model. In case of compression schemes without feedback the frame length has a big impact on the performance. The length of the frame has been chosen for each of the header compression scheme individually, such that the efficiency of the scheme is maximized. After initial simulations, the frame lengths are set to 0 packets. To apply Cooperative header compression, the existence of multiple paths between a source and a destination is required. For our studies we have chosen to work with Dynamic Source Routing Multipath Protocol (DSRMP) [] that is an extension to DSR protocol. The main objective of DSRMP is to provide multiple paths, both node-disjoint and link-disjoint, while keeping the signalling overhead low. DSRMP has only a slight increase in the routing overhead compared with DSR, but this is a price to pay for additional routes []. All simulation results presented in the paper are for DSRMP. Using compression schemes, from one hand, the delivery of packets is improved due to the smaller packet sizes and the fact that small packets are less prone to the channel

5 errors. From other hand, some packets can be lost due to decompression failure. Since we are interested in the combined performance of the routing and compression procedure, we consider a metric Packet Delivery and Decompression ratio. Figures and 5 show packet delivery and decompression ratio and bandwidth efficiency plotted versus node speed. For low node mobility compression results in the slight improvement in delivery ratio compared with the uncompressed case. But as the mobility increases, PDDR for delta coding scheme falls drastically: number of the lost packets increases due to route breakage and the decompression procedure fails often. This is a well-known draw-back of the schemes without the feedback. PDDR for scheme stays on the acceptable level - this can be explained by the performance gain due to the multipath diversity. From Fig. 5 we can see that applying header compression in ad hoc networks has a very positive effect on the bandwidth efficiency that can be also interpreted as achievable bandwidth savings. Fig.. PDDR Bandwidth efficiency uncompr delta coding Mobility (m/sec) Packet delivery and decompression ratio as a function of node speed uncompr delta coding Mobility (m/sec) Fig. 5. Bandwidth efficiency as a function of node speed. C. Overhead reduction The overhead reduction can be best represented by using the performance metric that is known in the field of IP header compression as compression gain. It is defined as the amount of information that is saved normalized to the summed size of headers if no header compression is used: G = N U N i= C i N U where N is a frame length, U is the size of uncompressed and compressed headers and C i represents a size of i-th compressed header in a frame. The average compression gain for delta coding and is found to be around 75% and 6% respectively. One should note that for different values of node speed there is some variations in compression gain due to the fact that the full context update is applied when a route is broken. Node mobility affects the number of route breakage and, thus, the amount of the uncompressed headers sent. V. CONCLUSIONS Header compression has a potential to become an attractive feature for ad hoc networks: reduction in packet size can increase the capacity and reduce congestion and delay. However, enabling header compression in ad hoc networks is more challenging than for other wireless networks due to unpredictable and highly dynamical ad hoc environment. We advocate the cross-layer approach and highlight the importance of the close interaction between the header compression algorithm and the routing protocol. REFERENCES [] R. Sridharan, R. Sridhar, S. Mishra, A Robust Header Compression Technique for Wireless Ad Hoc Networks, In Mobile Computing and Communications Review, Vol.7, No., 00. [] M. Kaddoura, S. Schneider, SEEHOC: Scalable and Robust End-to-End Header Compression Techniques for Wireless Ad Hoc Networks, In IEEE Globecom 00 Workshops, 00. [] F.H.P. Fitzek, S. Hendrata, P. Seeling, M. Reisslein. Chapter in Wireless Internet Header Compression Schemes for Wireless Internet Access. CRC Press. 00. [] A. J. Goldsmith, S. W. Wicker, Design Challenges for Energy-constrained Ad hoc Wireless Networks. IEEE Wireless Communications Magazine, Vol. 9(). 00. [5] M. Johansson, L. Xiao, Scheduling, Routing and POwer Allocation for Fairness in Wireless Networks. In IEEE VTC 00 Spring, 00. [6] Y.-C. Hu, D.B. Johnson, Implicit Source Routes for On-Demand Ad Hoc Network Routing, InProc. of the 00 ACM International Symposium on Mobile Ad Hoc Networking and Computing (MobiHoc 00), October, 00. [7] F.H.P. Fitzek, T. K. Madsen, P. Popovski, R. Prasad and M. Katz. Cooperative IP Header Compression for Parallel Channels in Wireless Meshed Networks. In Proc. IEEE International Conference on Communication (ICC), Seoul, Korea, 005. [8] T. Madsen, F. Fitzek, Y. Takatori, R. Prasad and M. Katz, Cooperative IP Header Compression using Multiple Access Points in G Wireless Networks, in Proceedings of IST Wireless Summit, 005. [9] J. Broch, D. B. Johnson, and D. A. Maltz. The Dynamic Source Routing Protocol for Mobile Ad Hoc Networks. Internet-Draft, draft-ietf-manetdsr-0.txt, October 999. [0] Network Simulator - ns-, [] A. Boursier, S. Dahlen, L. Marie-Francoise, T. S. Marin, S. Nethi, Multipath DSR Protocol for Ad hoc Networks, Report. Aalborg University. January 005.