A Self-Healing and Optimizing Routing Technique for Ad Hoc Networks

Transcription

1 Self-Healing and Optimizing Routing Technique for d Hoc Networks hao Gui and Prasant Mohapatra 1 bstract On demand routing protocols provide scalable and cost-effective solutions for packet routing in wireless ad hoc networks. The paths generated by these protocols may deviate far from the optimal because of the lack of knowledge about the global topology and the mobility of nodes. Routing optimality affects network performance, especially when the load is high. Longer route consumes more bandwidth, power, and are more prone to disconnections. In this paper, we propose a Self-Healing and Optimizing Routing Technique (SHORT) for mobile ad hoc networks that monitors the route and tries to shorten it if and when a shortcut path is available. Thus SHORT enhances performance in terms of bandwidth and latency without incurring any significant additional cost. irst, we have analytically derived the probability of existence of shorter paths, justifying the need for SHORT. Second, we propose a family of SHORT algorithms for different types of routing protocols. Third, we evaluate SHORT using the ns simulator. The results demonstrate that the performance of existing routing schemes can be significantly improved using the proposed SHORT algorithms. Index Terms d hoc Networks, Hop omparison rray, Routing Protocol, Self-Healing and Optimizing Routing Technique, Short-cut paths. I. INTROUTION Mobile d hoc Networks (MNTs) form a class of dynamic multi-hop network consisting of a set of mobile nodes that intercommunicate on shared wireless channels. ach node in MNTs can operate as a host as well as a router. Since the mobile nodes have limited transmission capacity, distant nodes intercommunicate through multihop paths. The ease of deployment and the lack of the need of any infrastructure makes MNTs an attractive choice for a variety of applications. xamples of such applications include communications in battle grounds, in disaster recovery areas, among a group of islands or ships, conferencing without the support of any wired infrastructure, and for interactive information sharing.. Gui and P. Mohapatra are with epartment of omputer Science, University of alifornia, avis, mail: This research was supported in part by the National Science oundation under the grants R and NI ontinuously changing network topology, low transmission power, and low available bandwidth are major challenges for routing in MNTs. Routing protocols should adjust to network topology changes, yet should not incur too much overheads in terms of the transmission of control messages. Normal link state routing, established in wired Internet routing, has failed to fulfill the special requirements of ad hoc routing [17]. This challenge has instigated intense research on routing in ad hoc networks in recent years. survey of several routing protocols and their performance comparisons have been reported in [1] and [16]. Recent proliferation of mobile wireless devices has considerably increased the need of ad hoc networking environments that connect large number of mobile nodes. With the increase in the size and average route length, scalability becomes an issue for the current routing protocols. Table-driven routing protocols [14] that require periodic advertisement and global dissemination of connectivity information are not suitable for large networks [11]. On-demand routing protocols [16] are efficient for routing in large ad hoc networks because they maintain the routes that are currently needed, initiating a path discovery process whenever a route is needed for message transfer. Popular protocols that have used this approach include OV [15] and SR [10]. In OV, the routing table at the nodes cache the next hop router information for a destination and use it as long as the next hop router remains active (originates or relays at least one packet for that destination within a specified time-out period). In SR, which is a source-based routing, identity of all the intermediate nodes are included in the packet header. In large ad hoc networks, the header length could become very long. ecause of the mobility of the nodes in ad hoc networks, the shape of routing paths may change significantly while the connectivity is intact. Most of the previously proposed on-demand routing schemes do not initiate a new path discovery process until there is a link failure (a node fails or moves out of range). The changes in shape can be exploited in deriving better routing paths if we can avoid any significant overheads (at least avoid extra path

2 2 discovery processes). In Section II, we show some examples of such scenarios and have discussed the potential performance impact of these changes in the shape of the path. Ill-formed paths that are much longer than the shortest available paths are not desirable. s data packets pass along these paths, extra bandwidth is consumed. Longer end-to-end delay and shorter route life-time are expected as well. In addition, since the number of hops in a longer path is high, the power consumption due to higher number of broadcasts is also high. mechanism that can optimize the route length will result in significant performance gain over the underlying routing protocols. In this paper, we have proposed a Self-Healing and Optimizing Routing Technique (SHORT) for mobile ad hoc networks that exploits the changes in shape of the routing paths because of the mobility of the nodes. The proposed schemes monitor the routing path and try to shorten the path length as and when it is feasible. shorter path not only reduces the latency but also enhances the reliability. Using the ns simulator we have shown the performance benefits of SHORT. It is observed that by employing SHORT on top of any routing algorithm results in better performance in terms of path optimality, delivery rate, and end-to-end delay with minimal overheads. Note that the benefit from SHORT is obtained when the path shape changes while the connectivity is intact. If the connectivity is broken then the path is determined through the path discovery process used in the underlying routing algorithm. The rest of the paper is organized as follows. The problem is stated in Section II with further description on scenarios in which short-cuts occur. Section III analytically estimates the probability of short-cut occurrence by using a geometrical probability model. Section IV illustrates the algorithm operations in detail, while Section V provides performance evaluation via simulations on ns-2 simulator. Related work are presented in Section VI, followed by the concluding remarks in Section VII II. PROLM SRIPTION onsider a routing path from a source node to a destination node I as shown in igure 1(a). This initial path is determined through the path discovery process, in which the distance between the source and destination is the shortest or very close to it. packet takes eight hops while getting routed from to I. uring the course of time, the mobility of the nodes may make the shape of the routing path similar to the one shown in igure 1(b) while retaining the connectivity. In this new shape, J is in the transmission range of, and is in the transmission range of J. Similarly H is in the transmission range of. However, because of the usage of route caches and the validity of the existing routing information, the routing table entries are not updated. lthough functionally adequate, using the routing paths of igure 1(b), a packet still takes eight hops to reach from node to node I. Ideally, the shortest path from to H needs only five hops as shown in igure 1(c). The goal of this paper is to identify such situations and self-heal and optimize the paths dynamically by modifying the entries of the routing tables. ig. 1. J J n example of the changes in routing paths. (a) (b) (c) The primary goal of the solution approach is to discover short-cut routing paths as and when feasible. The basic scenarios of the short-cut discovery process are shown in igure 2. In igure 2(a), the path -- can be reduced to - as is within the transmission range of. This short-cut path formation is termed as (2,1) reduction. In general, an ( ) reduction implies that routing hops are reduced to only one hop. igure 2(b) shows that the routing path --- can be shortened to --, since is in the range of, and is in the range of. This short-cut path formation is termed as (3,2) reduction. Thus, ( ) reduction implies that hops along the path can be reduced to only two hops. In general terms, ( ) reduction implies that routing hops can be reduced to hops, where. To avoid complexity in terms of overheads, we have considered the values of as 1 and 2 for the proposed SHORT algorithms. Intuitively, higher the difference between and, the better will be the performance of SHORT. This intuition is further validated in Section V. SHORT is applicable in conjunction with any underlying ad hoc routing protocol that formulates the entries of the routing tables. The underlying routing protocol need not have to be very efficient or optimal. It could be very simple and adequate enough to ensure a reachable path from a source to the destination. SHORT can self-heal the H H G G I I H I

3 & & 3 Transmision range a) Hop (2 1) short cut ig. 2. Utilized link Short cut link asic short-cut path formations. b) Hop (3 2) short cut ata packet Hyperthetical packets if short cut is taken path for optimization. The applicability of SHORT for distance vector type routing techniques is discussed in detail in section IV-. The algorithm can be also used for source-based routing technique like SR as is described in Section IV-. lthough ( ) short-cut is already supported in the original SR specification, we show the impact of SHORT algorithm on SR by using it with ( ) short-cut support, and the resulting performance is presented in Section V. III. PROILISTI MOL efore we describe the SHORT algorithm, it is necessary to show that the scenarios described in the previous section in fact occur with considerable probability. In this section, we derive the probability of the existence of the short-cut paths in mobile ad hoc networks. We model an ad hoc network by a set of random points on a twodimensional area. N points are plotted on an X-Y artesian plane, so that their x-coordinates and y-coordinates are uniformly distributed in the ranges of [0,X] and [0,Y], respectively. If the distance between any two points is less than 1 then they are regarded as adjacent. mong all these points, a sequence of given points are of interest, which are labeled as. We are given the condition that these nodes form a path, i.e. points and are adjacent for from to. This condition limits the relative positions of with respect to, since we know is located within a unit circle centered at. ut each point in the circle is equally probable to be the position of, so there is still a room for freedom for the shape of the path, and for the distance between and. or example, if the consecutive points form a straight line, then the distance can approach. On the other hand, if they form a cycle, the distance can be as low as zero. The question now is: What is the probability of and being adjacent. We denote this probability by, and term it as the probability of adjacency. Let s consider the two basic short-cut configurations shown in igure 2. Using the notations in our geometric model, the probability of hop (2,1) short-cut is just in adjacency probability. or a hop (3,2) short-cut, the adjacency probability is equal to, which is illustrated in igure 2(b). Nodes,,,, and form a sequence of five points, which can be denoted as,,,, and, respectively. ig. 3. r r+ r r/2 v V2 V1 P r 1 = djacency probability of three points. P 2 rea of shaded cycle band rea of unit circle r rea of shaded intersection = rea of unit circle = = 2*(rea of Right half) rea of unit circle 2*(rea of an rea of triangle) rea of unit circle or deriving, let us refer to igure 3. Let "! denote the probability that # is within the shaded cycle band centered at $ (i.e. %! '& %()*(,+.- / 0-+1(23'(0.) Let %! denote the probability of being within both s and s adjacency range. The geometrical derivations of! and 4! are shown in igure 3. The value of( is within the range 65 3,(. We thus have the following equations:? < & 798;:!>=!!! (1) #G/H HJI$K! 5!#L M5 Q! NPO 0(R ( (2) S5T$U V Q WYXZ []\$^ (3) Here, we have assumed that the mobile nodes are uniformly distributed within the given area. The node distribution in reality may differ from this to certain extent. However, this result should still be valid in providing an estimation on how frequently short-cut will form. To derive values of for larger, we define the following terms. Let _ & - J -`+am+ O be a series of random variables. or each one in the series, we define its probability distribution function b c( O as the following. b c( O & 7989: < 4()*(,+.- J -+d(se3,( 3'(! = Since any point within the unit circle centered at is equally probable to be the position of f, we can derive

4 4 TL I XPRIMNTL RSULTS OR VLUS O P[N]. ttempts , , the following for b c( O. b c( O & 0( +e('+ e( Referring to igure 4, we get the recurrence relation of distance distribution function: b c( O & v 0 b? 1( 5 0( HJI$K O b O R $R r radius=n Vn t θ Vn+1 radius=1 Under careful observations, it would be noted that the probability computed and simulated in this section, although correct, do not necessarily mean that these inefficient paths exist with such high probability. The computed probabilities include the cases of intermittent path breakages. However, in MNTs, when a path breaks, a new one is discovered. Nevertheless, the analytical and simulated results in this section do emphasize the point that we are not investigating isolated scenarios. These scenarios do occur frequently. Results in Section V further validates this point. IV. SHORT: TIL SRIPTIONS SHORT is a general technique that should work with any underlying routing protocol. ue to the lack of precise global information and distributed operational manner, we should not expect routing protocols to make globally optimal decisions at the route setup time. However, as long as a valid routing path is set up, and packets begin to flow along the path, we will let SHORT monitor the path. It will try to identify short-cuts on the route and shorten it according to the up-to-date topology. ifferent types of routing protocols work in different manner, we need to have different algorithms for short-cut identification and reaction. SHORT-V algorithm works with distance vector routing protocols such as OV, while SHORT-SR algorithm works with source routing protocols such as SR. It should be noted that the SHORT algorithm does not react to every little change in the topology or for transient conditions. This issue is addressed in section IV-. ig. 4. Recurrence relation of. Using the distance distribution function, we can gain the probability of adjacency: & b c( O R ( [ To obtain empirical values for, we conduct random walk experiments simulated by computer. ach experiment starts from the position of and begin taking steps. or each step, the target point is equally likely to be any position as long as the step length is less than 1. fter taking steps, if the final position is still within the reach of one step range, we call the experiment successful. We repeat each experiment 10,000 and 100,000 times. Table I gives the frequency of success for different values of. It shows that the probability of (3,2) short-cut is about [. The simulation results also validates our analytical method.. SHORT-V lgorithm When a node wants to forward a packet to another node, it broadcasts the packet and all the nodes within the transmission range can hear the packet. ach of the nodes check the packet header to see if the packet is destined to it (as the next hop) or not. The nodes that is the next hop destination, captures the packet and processes it for the next hop. Other nodes disregard the packet (unless they are operating in promiscuous mode). So in any case, the neighboring nodes (nodes that are within the transmission range) get to check all the packet headers that they hear. To support the proposed SHORT algorithm, each packet needs to carry a hop-count (H) field in its header. The H field is initialized to zero at the source node and gets incremented by one at every hop the packet takes. or every packet, the destination address (), source address (S), and the H can be obtained from the packet header. This information is maintained as an array termed as the hop comparison array. The format of

5 5 each entry in the array is, where is the neighbor s address from which the packet was broadcast. ach of the element of the array has an expiration time after which they are invalidated. If the array becomes full with valid entries, the new entries can replace the older entries using the least-recently-used (LRU) policy. onsider a source node and a destination node. efore sends a packet to the first hop node, an entry with the following information is included in s hop comparison array: >. or an incoming packet at a node, the steps of the SHORT algorithm can be enumerated as shown in lgorithm 1. lgorithm 1 SHORT-V lgorithm I the node is the final destination address, consume the packet. LS I the packet is destined to as the next-hop node, THN process the packet for forwarding further. I an entry corresponding to the < > does not exist in the hop count array, THN record the information: < >. LS ompare < > to all the valid entries in the hop comparison array. I there is no match with the entries, THN store < > in the hop comparison array. LS (ssume that it matched with an entry < >) 5 I, THN Send a message to to update the routing table such that the next hop address for destination node is modified to the address of node. The node modifies its routing table by making the next-hop address for destination as. The entry corresponding to < > is deleted from the hop comparison array maintained at node. N lgorithm 1 can be explained as follows. t the final destination, a packet is always consumed without any events. t a node, which is the next-hop for a broadcast is recorded in the hop packet, comparison array if no entry exist corresponding to the! pair. If an entry for! exist then nothing is updated. The comparison of Hs and the modification of the routing table (when necessary) happens only at the other nodes, i.e., excluding the current node and the next-hop node. In those nodes, if an entry corresponding to " # does not exist then it is recorded. Otherwise the current H is compared with the stored H. If the difference (current H stored H) is more than 2 then it implies that a short-cut path exists. The value of 2 is used as we can shorten the path of a minimum of 2 hops. The node that detects this possibility modifies its routing table entry and that of another appropriate node to enable the short-cut path formation. Note that the lgorithm 1 is capable of handling both the (,1) and (,2) short-cut configuration discussed earlier in Section II. The proposed SHORT algorithm does not incur any significant overhead. ach of the nodes use a very small amount of memory space to maintain the hop comparison array. or every short-cut path formation, only one extra broadcast message (for one time only) is necessary to inform one of the neighboring nodes to update its routing table. 1) Tracing an xample: Let us analyze an example to show the functionalities and validate the correctness of the algorithm. onsider the routing path shown in igure 1(b). We will trace through the algorithm to analyze how the routing path self-optimizes itself without any significant overhead. ssume node as the source and node I as the destination. The current routing path is shown by the solid lines in the figure. We will track only one entry of the hop comparison array maintained at each of the nodes. This entry corresponds to (,I) as the sourcedestination pair. The example is analyzed in steps where a step defines the events during the broadcast of a packet at a node. The entry of the hop comparison array at end of each step at each of the node is shown in the Table II. i needs to forward a packet to. n entry is initialized in s hop comparison array as (,I,0,). It broadcasts the packet while marking the H as 0. and J are in the transmission range and receive the broadcast. is the next-hop node and does not have any entries corresponding to (,I). So it records (,I,0,) in its hop comparison array. J also did not have any entry corresponding to (,I), so it records (,I,0,) in its hop comparison array. ii To forward the packet to, broadcasts the packet

6 6 TL II TH NTRY O TH HOP OMPRISON RRY. G H I J Short-cut i (0,) (0,) (0,) ii (0,) (0,) (1,) (0,) iii (0,) (0,) (1,) (2,) (0,) iv (0,) (0,) (1,) (2,) (3,) (0,) v (0,) (0,) (1,) (2,) (3,) (4,) (4,) ound at J vi (0,) (0,) (1,) (2,) (3,) (4,) (5,) (5,) vii (0,) (0,) (1,) (2,) (3,) (4,) (5,) (5,) viii (0,) (0,) (1,) (2,) (3,) (7,H) (5,) (5,) ound at with H=1. Only and are in the transmission range. records (,I,1,) in its hop comparison array. finds a match with its entry in the hop comparison array. However, the difference in the Hs is not more than 2, so nothing is updated. iii broadcasts the packet to forward it to with an H of 2. Only and are in range. records (,I,2,), and nothing is updated at since the difference in Hs is 2 (needs to react only when it is more than 2). iv Similarly when forwards the packet with only and in the range, the entry of the hop comparison array at is noted as (,I,3,), and everything else the same. v When broadcasts the packet to send it to, then,, and J are in the transmission range. Nothing is updated at, and the entry of the hop comparison array at is recorded as (,I,4,). The H difference at node J is 4. So J sends a message to to update its routing table so that the next hop of for destination I is J. J modifies its own routing table to make the next-hop node for destination I as. Thus the path -J- is created as a short-cut for the earlier route The corresponding entries at nodes, J, and are deleted (the reason for the deletion is explained later). vi broadcasts the packet with a H of 5 destined for G. Nodes and H also receive the packet as they are in the transmission range., G, and H record (,I,5,) in their hop comparison array. vii G broadcasts the packet destined to H with a H of 6. oth the receivers in the range, and H, have an entry corresponding to (,I). However, since the difference in Hs is not more than 2, nothing is updated. viii While forwarding the packet to I, H broadcasts the packet with a H of 7. I, G, and are in the transmission range. Node I consumes the packet as it is the destination node. G compares the H with its stored entry. s the difference in H is not more than 2, nothing is updated at G. t the H difference is 3. So sends a message to to update its routing table to point to for destination I. In this case, the routing table entry already points to, so no update is necessary. This step is redundant in this case. However, it maintains the correctness of the algorithm. then updates its own routing table to point to H as the next-hop node for destination I. The corresponding entry of the hop comparison array at,, and H are deleted. Thus the path -G-H is shortened to -H. ix The final path that is obtained is as shown in igure 1(c). 2) Special ases: There are some special cases in the topologies of ad hoc network which are not exploited by the basic SHORT algorithm. dditional tuning of the algorithm is necessary to capture these occurrences. However, since these occurrences would be isolated, we do not modify the algorithm to make it more complex to handle just a few cases. In any case, the algorithm remains functionally correct in such situations, although not optimized. ig. 5. SOUR (2,1) short-cut involving the source node. onsider the part of an ad hoc network shown in igure 5. When the source node broadcasts a message intended to some destination Q, it stores (,Q,0,) in its hop comparison array. oth and, which are in the trans-

7 7 mission range of, record the same information in their respective hop count arrays. Nothing is updated while forwards the packet to. When forwards the packet to, the H is 2. When this H is compared with the H stored at, the difference is not more than 2. So the shortcut route - is not detected. Note that this happens only when is the source node. If would have been a transit node, then the H difference should have been more than 2. Thus the short-cut process of (2,1) is not detected by the source node. However, all other short-cuts (,1) or (,2) are very well detected by the source node when is more than 2. nother special case where the SHORT algorithm fails to discover the short-cut path involves the destination node. These isolated configurations are shown in igure 6. Since the destination nodes consumes the packet without any further broadcast, the short-cut paths are not detectable by the nodes (in igure 6(a)) or node G (in igure 6(b)). If necessary, these isolated events can be subsumed by the proposed algorithm by enforcing the destination node to do an additional broadcast of the packetheader while consuming the packet. In that case, the node or G would be able to detect and create the short-cut paths. (a) STINTION packet routing. TL III TH NTRY O TH HOP OMPRISON RRY. i (0,) (0,) ii (0,) (0,) (1,) (1,) iii (0,) (0,) (1,) (1,) (2,) iv (0,) (0,) (1,) (1,) (2,) v (0,) (0,) (1,) (1,) (2,) Using the Table III, we can observe that when broadcasts the packet to forward it to, then creates the short-cut path -- instead of --- as shown in igure 7(b). It becomes very interesting when forwards the packet to the destination. t that point, both and discover that a short-cut route can be created. So either of them create their own short-cuts. If creates the short-cut path first and creates later, then the final configuration would be the topology shown in igure 7(b). However, if creates the short-cut path first and then creates the short-cut, then the final path would be the topology shown in igure 7(c). In either case, the final path provides the connectivity as well as minimizes the length of the path from to. (a) (b) G STINTION ig. 6. (b) Short-cut involving the destination node. ig. 7. (c) Short-cut formation for zig-zag paths. n interesting observation can be made in circumstances where the topology of the network would form a zig-zag fashion as shown in igure 7(a). Let be the source and be the destination. The relevant entry of the hop count array is shown in Table III for each step of the. SHORT-SR algorithm In source initiated routing techniques, such as SR [10], the entire path (i.e., the address of all the nodes that the packet will hop through) is encoded in the packet

8 8 header. In [10], the authors have already shown how the (,1) short-cut situations are handled. SHORT can be also employed on top of SR to facilitate path formation with (,2) and higher short-cuts in addition to the (,1) shortcuts. ccording to the original specification of SR protocol[2], [10], (,1) short-cut is discovered as follows. ach node is in promiscuous listening mode. When node taps a packet sent out by node, but is not the intended nexthop receiver, node checks the route given in the packet to see if it is in the up-coming part of the list of intermediate nodes on the route. If is on the list and it is more than 2 hops away from on the routing path, finds a short-cut. then sends a gratuitous RRP message to the source node of the route informing that to part of the route can be reduced to link -. The gratuitous RRP message travels back along the shortened path in reverse direction. (,2) short-cut can be discovered in a similar way. We assume that each node can maintain a current neighbor table from the tapped packets. When node taps a packet sent out by node, node put into its neighbor table. Then checks the route given in the packet to see if any of its neighbors, except, is in the up-coming part of the routing path. If node, a node in s neighbor table, is in the path more than 3 hops away from, finds a shortcut. then send out the gratuitous RRP message.. iscussion on Implementation When implementing SHORT-V algorithm, we find a need to control the aggressiveness of the algorithm. Without this curbing, short-cut messages interfere with the routing tables, causing much more route updates than the original protocols. This problem is caused by ephemeral short-cuts and multiple short-cuts. phemeral short-cut happens when a fast-moving node finds out itself in a position to form a short-cut and informs the involved on-route nodes, but actually it only can stay in the effective shortcut position range for a short time. Multiple short-cuts means that more than one node stays within the position range of one certain short-cut. So they all report to the same involved on-route nodes, which causes the informed node to change the relevant routing table entry too frequently. However, responding only one short-cut report is enough. or a node to identify ephemeral short-cut, it needs to know its position and moving speed. Such support may not be available for the ad hoc network. Multiple short-cut problem cannot be prevented without introducing extra overhead of message exchanging. In other words, it is not easy to control the credit of short-cut messages. We can resort to the receiver side to solve this problem, by making each node conservative in accepting short-cuts. We introduce the concept of a stable period for any newly entered or updated entry in routing table, so that each node ignores any update to it within the stable period. In this way, we can protect the consistency of routing tables. The fine tuning of this parameter and their consequences are discussed in the following section. Source routing protocols are more immune to interferences from short-cuts. In SHORT-SR, each short-cut takes the form of a new route. The new route is sent back to the source node, which puts it into the node s route cache. So multiple short-cuts do not impose problems for SHORT- SR, since this only leads to the growth of route cache at the source node. However, ephemeral short-cuts are not tolerable. If a short-cut becomes invalid before the informing message gets to the source node, the route cache at the source node is polluted with dirty information. We address this problem by assuming temporal locality[4], which means the longer a node has stayed within a valid short-cut position range the more is the probability that it can form a stable short-cut path.. Simulation Setup V. PRORMN VLUTION Our simulation is based on ns-2 developed by the VINT project 1, and further extended by Monarch project 2 to support wireless network. The simulator provides a proper model for signal propagation, and the radio model is based on Lucent Technologies WaveLN product, with 2Mbps of transmission rate and 250 meter of transmission range. The I is simulated at the M layer, with implementation of the distributed coordination function(). In this simulation we let 50 mobile nodes moving within a rectangular field of X in size. We choose this rectangular field so that the average hop distance between any two nodes will be larger than that of a square field with same area. The duration of each run is 900 simulated seconds. The mobility model uses the random waypoint model[8]. We change mobility rate by setting different values to pause time as 0, 45, 90, 180, 270, 540, 720 and 900 simulated seconds. Here a pause time of zero means continuous mobility, and 900 seconds reflects a scenario of stable nodes. The maximum moving speed can be 10m/s or 20m/s. We ran simulations covering each combination of pause time and moving speed. or the traffic model, we use 20 simultaneous sessions with source-destination pairs spreading randomly on The URL to their website is N The URL to their website is

9 9 Routing Optimality(Length of actual route / Shortest distance) ig OV OV-SHORT SR SR-SHORT 1.02 omparison of routing optimality elivery Rate ig OV OV-SHORT SR SR-SHORT 0.88 omparison of packet delivery rate. the whole network. Traffic sources are constant-bit-rate, sending 4 UP packets a second. ach packet is 512 bytes long, thus resulting a 2K byte per second data transfer rate for each session. TL IV ONSTNTS US IN OV-R onstant Value TTL STRT 1 TTL INRS 2 TTL THRSHOL 7 We implemented both SHORT-V algorithm on OV protocol, and SHORT-SR algorithm on SR protocol. The enhanced version of the both protocols are referred to as OV-SHORT and SR-SHORT, respectively. or both original OV and OV-SHORT, expanded ring search(r) technique is used. or each route discovery, first RRQ packet has a small time-to-live(ttl) value given by TTL STRT constant. In every subsequent RRQ packet, the TTL field is increased by a value according to TTL INRS. When that value reaches a TTL THRSHOL, a network wide flooding is finally used. The actual values used in the simulation are shown in Table IV.. Results and evaluations 1) Path Optimality: We define path optimality as the ratio of the number of hops a packet took to reach its destination over the shortest hop distance between the sourcedestination pair at the time the packet is sent. We average this value over all received packets during a simulation run. or an ideal case when every packet takes the shortest path, this metric will be 1. In general, a value more closer to 1 indicates better routing optimality for the routing protocol. igure 8 shows the averaged path optimality as a function with respect to pause time. Starting from no mobility scenario to higher mobility rate, OV significantly drops in its routing optimality. or the continuous movement scenario, at an average, every packet travels a route 18% longer than the shortest paths. This problem is corrected by SHORT algorithm. or all pause time values, the change in routing optimality is not significant. Note that even at no mobility scenario, OV- SHORT still achieves better optimality than OV. This is because OV does not guarantee shortest routing path even when all nodes are stable, but during the simulation run, SHORT will gradually heal and rectify if there is a short-cut, thus resulting in better routing optimality. SR, on the other hand, achieves better performance in path optimality than OV. This is because it already enjoys (,1) short-cut support, which also explains that SR has the same performance with SR-SHORT for stable (no movement) scenario. We notice that in general SR- SHORT does not achieve as good a gain of path optimality as OV-SHORT. This is because in SR-SHORT, a short-cut will not take into effect until the informing message gets to the source node, which is a relatively long delay that does not exist in OV-SHORT. 2) elivery rate: Packet delivery ratio of a session is the ratio of number of packets that are received by the destination over the number of packets submitted to the network by the R source. We look at the overall delivery ratio averaged over all 20 sessions. igure 9 shows the overall delivery ratio as a function with respect to pause time. t low mobility rate, there is not much difference when SHORT is added. s the mobility rate increases, OV s performance drops faster than SR, which is reflected by the significant drop of path optimality in OV. In this simulation, the local repair option is turned off in the OV implementation. It will achieve better delivery rate with the option turned on, but the path optimality will become worse[11]. oth OV-SHORT and SR-SHORT achieves a stable performance at all mobil-

10 OV SHORT-0.1s SHORT-0.5s SHORT-1.0s OV SHORT-0.1s SHORT-0.5s SHORT-1.0s Routing Overhead(Number of routing packets) Number of Route iscoveries (a) Number of routing packets sent. 0 (b) Number of route discoveries. ig. 10. omparison of Routing Overhead on OV and OV-SHORT with different values of route stable time. Routing Overhead(Number of routing packets) SR-10m/s SR-SORT-10m/s SR-20m/s SR-SORT-20m/s Number of Route iscoveries SR-10m/s SR-SHORT-10m/s SR-20m/s SR-SHORT-20m/s (a) Number of routing packets sent. 0 (b) Number of route discoveries. ig. 11. omparison of Routing Overhead on SR and SR-SHORT with maximum moving speed as 10 m/s and 20 m/s. ity rates. ven at continuous mobile scenario, i.e. pause time is 0, more than 97% of the packets are delivered. This result conforms with the analysis in previous sections, that the loss rate is proportional to the route length as the loss probability increases with each additional hop. 3) Routing Overhead: We look at routing overhead in two aspects: the total number of routing packets transmitted and the total number of route request packets initiated during the simulation. or each routing packets sent over the multi-hop path, each hop is counted as one transmission. or the number of route requests, we only count its initiation at source nodes, and its re-broadcast at other nodes is not counted. or SR-SHORT and OV- SHORT, routing overhead also includes short-cut informing messages as well. igure 10 shows the comparison of routing overhead on OV and OV-SHORT. The maximum moving speed is 20 m/s. With SHORT algorithm, we set the route stable time as 0.1, 0.5 and 1.0 seconds (and they are referred as SHORT-0.1, SHORT-0.5 and SHORT-1.0 respectively). These figures show that routing overhead changes dramatically with route stable time when mobility ratio is high. SHORT-0.1 performs badly at high mobility rate. It heavily interferes with the correctness of OV routing, initiating much more route requests than other protocols. s the the mobility rate becomes lower, the influence of SHORT algorithm decreases, until it becomes nearly zero when the mobility is zero. SHORT-1.0 costs a similar overhead with original OV. It always requires less number of route discoveries than OV, but its overhead is a little higher in high mobility. This extra overhead

11 11 comes from short-cut informing messages. ut when mobility becomes lower, SHORT-1.0 out-performs original OV. SHORT-0.5 has the best performance. Its overhead is stable with respect to the change in mobility rate. igure 11 shows the comparison of routing overhead on SR and SR-SHORT. There is no route stable time in SR-SHORT. We show two scenarios in the figure: with 10 m/s and 20 m/s maximum moving speeds. The performance improvements is not as much as in OV. or both scenarios, from zero to 200 seconds in pause time, SR-SHORT sends out more routing packets than SR. Then, at longer pause times, SR-SHORT begins to send less routing packets than SR. s for the number of route requests initiated, SR-SHORT is always less than SR in all pause times. So (,2) short-cuts do help in reducing route breakages. SR-SHORT incurs less overhead in relaying RRQ and RRP packets than SR. However, at high mobility cases, short-cut informing messages proliferates, and they are relayed to the source nodes. This makes SR-SHORT s overall routing overhead higher than SR. 4) nd-to-end elay: Next, we compare the end-toend delay benefits obtained through the usage of SHORT. This metric is also averaged over all received packets, and it is plotted in igure 12. irst, we should note that this comparison favors the protocols with lower delivery rate. The reason is that only the packets that are actually received are counted, and these packets tend to travel in shorter route, thus have smaller end-to-end delay. ven with this conservative estimate, a significant improvement is obtained by using SHORT algorithm, which is due to the difference of routing optimality and number of route breakages. Here, the implementation of OV and SR do not drop the suffered packets when the relevant link breaks, but put them in a queue and resend them when a new route is eventually setup or overheard from other nodes. VI. RLT WORK In source routing schemes such as SR, intermediate on-route nodes have information about the entire route. aching of these information provides nodes with the knowledge of the topology of nearby nodes[8]. Thus routes can be found within this topology and (,1) shortcuts can be easily found as well. ut (,2) short-cuts are not supported, and this caching scheme cannot work with other protocols like OV [15] or TOR [13]. Neighborhood awareness is exploited in neighborhood-aware source routing(nsr)[19]. Topology of nodes within two hops are learned from the source routing protocol. This nd-to-end elay (milli-second) ig. 12. packets OV OV-SHORT SR SR-SHORT 0 omparison of end-to-end delay averaged over all received information is used to repair a broken route by replacing the broken link with an alternative short route. OV-R[12] has some similarity to SHORT in the sense that it lets neighboring nodes around a route be aware of the route. ach neighboring node computes an alternative short route for a nearby link on the route, and when the link breaks, relevant neighboring node will salvage the route and help to deliver the affected packets to the destination. Several reports have proposed techniques toward improving the route stability. ssociativity-ased Routing (R)[21] protocol defines a routing metric known as degree of association stability. high value of the metric may indicate a low state of relative mobility between neighboring nodes. In R, a route is selected based on the degree of association stability of mobile nodes, so that it can derive longer-lived routes. similar approach is used in Signal Stability-ased daptive Routing(SSR)[5], in which route selection is based on the signal strength between nodes and a node s location stability. Routes that have stronger connectivity are preferred, which may also result in more stable route. Mobility prediction [20] is a different technique toward enhancing route stability. Utilizing the position and velocity information provided by GPS support, the expected expiration time of each link can be calculated, and the future state of network topology can be predicted. ased on this prediction, route can be reconstructed before it breaks, so that the disruptions caused by the changing topology are minimized. Scalability issues in ad hoc networks routing are studied in [18] and [11]. In [18], a unified analytical performance model is developed for different routing protocols, reflecting the impact of network size, traffic intensity and mobility on protocol performance. whole chapter in [11] is devoted to scalability issues. Results of simulations on large size networks with 10,000 nodes have shown that the average length of paths generated by on-demand routing

12 12 protocols is more than double the optimal length. This causes performance degradations in terms of throughput and end-to-end delay. VII. ONLUSION In this paper, we proposed a self-healing and optimizing routing technique for mobile ad hoc networks. SHORT improves routing optimality by gradually shortening the routes whenever possible. The basic idea is to let neighboring nodes of a routing path, together with the on-route nodes, monitor the route, so that up-to-date relative topology information is exploited in identifying shortcuts. When short-cut occurs, and is estimated to be stable, it will be utilized to shorten the route. We have analyzed and evaluated this technique for both OV and SR algorithms, using the ns simulator. Simulation results show that higher delivery rate and lower end-to-end delay is achieved. Up to now, in self-optimizing operations, we only use route length as optimization metric. We can use other metrics as well, such as QoS requirements and other state information. This will be the theme of our future work. RRNS [1] J. roch,.. Maltz,.. Johnson, Y.-. Hu, and J. Jetcheva, Performance omparison of Multi-Hop Wireless d Hoc Network Routing Protocols, Proc. M International onference on Mobile omputing and Networking (MOIOM), [2] J. roch,.. Johnson and.. Maltz, The ynamic Source Routing Protocol for Mobile d Hoc Networks, IT Internet draft, draft-ietf-manet-dsr-01.txt, ec (work in progress.) [3] R. astaneda and S. R. as, Query Localization Techniques for On-demand Routing Protocols in d Hoc Networks, Proc. M International onference on Mobile omputing and Networking (MOIOM), [4] S. hen and K. Nahrstedt, istributed Quality-of-Service Routing in d-hoc Networks, I Journal of Selected reas on ommunications, vol. 17, No. 8, ugust [5] R. ube,.. Rais, K.-Y. Wang, and S. K. Tripathi, Signal Stability based daptive Routing (SS) for d-hoc Mobile Networks, I Personal ommunications, pp , ebruary [6] K. all and K. Varadhan (ds.), ns notes and documentation, 2002, available from [7]. Ganesan, R. Govindan, S. Shenker and. strin, Highly- Resilient, nergy-fficient Multipath Routing in Wireless Sensor Networks, Proc. M International Symposium on Mobile d Hoc Networking & omputing (MOIHO), [8] Y.-. Hu and.. Johnson, aching Strategies in On-emand Routing Protocols for Wireless d Hoc Networks, Proc. M International onference on Mobile omputing and Network (MOIOM), [9]. Iwata,.. hiang, G. Pei, M. Gerla and T. hen, Scalable Routing Strategies for d Hoc Wireless Networks, I Journal of Selected reas on ommunications, vol. 17, No. 8, ug [10].. Johnson and.. Maltz, ynamic Source Routing in d-hoc Wireless Networks, Mobile omputing, T. Imielinski and H. Korth, ds., Kluwer, 1996, pp [11] S. J. Lee, Routing and Multicasting Strategies in Wireless Mobile d Hoc Networks, Ph.. Thesis, 2000, omputer Science epartment, UL. [12] S. J. Lee and M. Gerla, OV-R: ackup Routing in d Hoc Networks, Proc. I WN 2000, hicago, IL, Sep [13] V. Park, S. orson, Highly daptive istributed Routing lgorithm for Mobile Wireless Networks, Proc. I INOOM 97, Vol. 3, pp [14].. Perkins and P. hagwat, Highly ynamic estination- Sequenced istance-vector Routing(SV) for Mobile omputers, omp. ommun. Rev., Oct. 1994, pp [15].. Perkins and. M. Royer, d-hoc On-emand istance Vector Routing, Proc. I Workshop on Mobile omputing Systems and pplications, pp , eb [16]. M. Royer and -K. Toh, Review of urrent Routing Protocols for d Hoc Mobile Wireless Networks, I Personal ommunications, pril [17].. Santivanez, R. Ramanathan and I. Stavrakakis, Making Link-State Routing Scale for d Hoc Networks, Proc. M International Symposium on Mobile d Hoc Networking & omputing (MOIHO), [18].. Santivanez,. Mconald, I. Stavrakakis, and R. Ramanathan, On the Scalability of d Hoc Routing Protocols, Proc. I INOOM 2002, New York, NY, June [19] M. Spohn and J.J. Garcia-Luna-ceves, Neighborhood ware Source Routing, Proc. M International Symposium on Mobile d Hoc Networking & omputing (MOIHO), [20] W. Su, S.-J. Lee and M. Gerla, Mobility Prediction in Wireless Networks, Proc. I MILOM 2000, Los ngeles,, Oct [21] -K. Toh, Novel istributed Routing Protocol To Support d-hoc Mobile omputing, Proc I 15th nnual International Phoenix onf. omp. and ommun., Mar. 1996, pp