1568 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 59, NO. 3, MARCH 2010 a path consisting of nodes that may survive for the longest time among lifetime of the route becomes the minimum lifetime of all links in this multiple paths. Shrestha and Mans  mentioned that the energy drain route. A link is formed by two adjacent mobile nodes, which have rate of a node is affected not only by its own but by its neighboring data limited battery energy and can roam freely, and it is broken if any ﬂows as well. Marbuhk and Subbarao  aimed to preserve network of the two nodes is not alive due to exhaustion of energy or if these connectivity by choosing a route according to the remaining battery two nodes move out of each other’s communication range. We use the life of nodes along the route. Toh  proposed selecting a path with connection lifetime and the node lifetime to distinguish the two cases minimum total transmission power when there exist some possible described above. The connection lifetime is described as LLT in – paths, and all nodes through these paths have sufﬁcient residual battery . In this paper, a link is composed of the two nodes in a connection power. Misra and Banerjee  proposed selecting a path that has and the connection itself, and the LLT includes both the node lifetime the largest packet transmission capacity (the residual energy divided and the connection lifetime. by the expected energy spent in reliably forwarding a packet) at a A link Li consists of a connection Ci and two nodes (Ni−1 , Ni ), “critical” node among multiple paths. The critical node is the node where Ci represents the connection between nodes Ni−1 and Ni , that has the smallest packet transmission capacity in a path. In the and it is maintained until the adjacent nodes (Ni−1 , Ni ) move out of lifetime-prediction routing (LPR) algorithm , each node attempts each other’s communication range under the assumption of no energy to estimate its battery lifetime based on its residual energy and its past problem in both nodes Ni−1 and Ni . We introduce connection lifetime activity. The above algorithms used well-deﬁned metrics to evaluate TCi to represent the estimated lifetime of the connection Ci , and it the lifetime of nodes. However, all of them ignored the mobility of only depends on their relative mobility and distance of nodes Ni−1 mobile hosts, and thus, it seems that they are more suitable for static and Ni at a given time. The term TNi denotes the estimated battery networks. lifetime of node Ni . Then, the lifetime of the link Li is expressed as The LLT routing algorithms are used to estimate the lifetime of the minimum value of (TCi , TNi−1 , TNi ), i.e., wireless links between every two adjacent nodes and then to select an optimal path. In the associativity-based routing algorithm , a link is TLi = min(TCi , TNi−1 , TNi ). (1) considered to be stable when its lifetime exceeds a speciﬁc threshold Take as an example a route P consisting of n links. Route P is said that depends on the relative speed of mobile hosts. In the signal- to be broken if any one of the following cases occurs. First, any one of stability-based adaptive (SSA) routing , each link is classiﬁed as the nodes in the route dies because of limited battery energy. Second, a strong one or a weak one, depending on the received signal strength any one of the connections is broken because the corresponding two measured when a node receives data packets from the corresponding adjacent nodes move out of each other’s communication range. Thus, upstream node. A mobile node only processes a route request (RREQ) the lifetime of route P is expressed as the minimum value of the that is received from a strong link. Tickoo et al.  computed the lifetime of both nodes and connections involved in route P. Assume fragility of a link as the difference of the received signal strengths of that Ω represents the set of all nodes in route P and that Ψ is the set of consecutive packets ﬂowing from the same origin to check if these two all the connections in route P. Thus, the lifetime Tp of route P can be nodes are getting closer or moving apart. Gerharz et al.  predicted expressed as the lifetime of a link between two adjacent mobile hosts through online statistical analysis of the observed links. Several studies – Tp = min (TNi , TCi ) . (2) attempted to predict the expiration time of the links by estimating Ni ∈Ω,Ci ∈Ψ the mobility of nodes. Note that all of these studies assumed for simplicity that mobile hosts are kept at a constant speed and direction From (2), the lifetime Tp of route P is estimated from the prediction in a short period. Samar and Wicker  developed an analytical lifetime of each node and each connection. framework to investigate the behavior of nodes in a random mobility environment and derived analytical expressions that are related to the A. Node Lifetime Prediction Algorithm lifetime of links. As a different approach, Wu et al.  used a two- state Markovian model to reﬂect the mobility of nodes and evaluate the Intuitively, if there are two nodes that have the same residual energy link dynamics. level, an active node that is used in many data-forwarding paths In MANETs, a route consists of multiple links in series, and thus, its consumes energy more quickly, and thus, it has a shorter lifetime than lifetime depends on the lifetime of each node, as well as the wireless the remaining inactive node. With the same concept as described in links between adjacent nodes. The main contribution of this paper is , we evaluate the node lifetime that is based on its current residual that we combine node lifetime and LLT in our route lifetime-prediction energy and its past activity; however, we present a much simpler algorithm, which explores the dynamic nature of mobile nodes (i.e., solution that does not need to calculate the predicted node lifetime the energy drain rate of nodes and the relative mobility estimation from each data packet. rate at which adjacent nodes move apart) in a route-discovery period The term Ei represents the current residual energy of node i, and that predicts the lifetime of routes discovered, and then, we select the evi is the rate of energy depletion. Ei can simply be obtained online longest lifetime route for persistent data forwarding when making a from a battery management instrument, and evi is the statistical value route decision. The rest of this paper is organized as follows. Section II that is obtained from recent history. We use an exponentially weighted describes the proposed route lifetime-prediction algorithm. Section III moving average method to estimate the energy drain rate evi . implements our proposed algorithm in an exploring dynamic nature Every T seconds (the smaller the value of T , the more accu- routing (EDNR) protocol environment based on dynamic source rout- rate this algorithm is), node i reads the instantaneous residual en- 0 2T 3T (n−1)T nT ing (DSR). Section IV presents the performance-evaluation results. ergy value Ei , Ei , Ei , . . . , Ei , Ei , . . ., in every period Finally, Section V draws the conclusion. [0, T], [T, 2T], [2T, 3T], . . . , [(n − 1)T, nT] . . ., and the correspond- ing estimated energy drain rate evi is obtained as ⎧ 0 II. ROUTE L IFETIME -P REDICTION A LGORITHM ⎪ evi = 0, ⎨ n=0 evi = Ei − Ei /T, 1 0 T n=1 (3) Since a route consists of multiple links in series, it is said to be ⎪ ev n = α E (n−1)T − E nT /T + (1 − α)ev n−1 , n > 1 ⎩ i broken if any single link among its links is broken, and thus, the i i iAuthorized licensed use limited to: University of Science and Technology of China. Downloaded on March 18,2010 at 04:40:01 EDT from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 59, NO. 3, MARCH 2010 1569 Fig. 1. Relative motion. Fig. 2. LLT prediction algorithm. n where eviis the estimated energy drain rate in the nth period, and evin−1 is the estimated energy drain rate in the previous (n − 1)th To calculate the connection time TCi , we apply a triangle geometry n period. α denotes the coefﬁcient that reﬂects the relation between evi theory and improve the method that needs three sample packets and n−1 predicts the link expiration time for proactive route maintenance in the and evi , and it is a constant value with a range of [0, 1]. In this algorithm, to reﬂect the current condition of node i well, we grant a previous work . Our proposed method requires only two sample n higher priority to evi and set α = 0.6. packets, and we implement piggyback information on route-request At time t, we can obtain the estimated node lifetime as follows: (RREQ) and route-reply (RREP) packets during a route-discovery procedure (details described in Section III) with no other control nT n TNi = Ei /evi , t ∈ [nT, (n + 1)T ] . (4) message overhead, and thus, it does not increase time complexity. Fig. 2 illustrates the proposed LLT prediction algorithm. If node Ni is set to the reference frame, node Ni−1 moves at velocity v relative B. Connection Lifetime-Prediction Algorithm to the velocity of node Ni . Ni−1 receives two packets from node Ni We evaluate the LLT using the connection lifetime; however, it is at time t0 and t1 . We assume that node Ni−1 moves out of node Ni ’s difﬁcult to predict the connection lifetime TCi between two nodes radio transmission range at prediction time t. (Ni−1 , Ni ) because the nodes in MANETs may move freely – At time t0 , node Ni−1 receives a packet from node Ni , and the . In our algorithm, we only handle the connections that are in an received signal power is P0 ; thus, the distance d0 between the two unstable state and may only last for a short period particularly, ignoring nodes can be calculated by using a radio-propagation model . the stable one for simplicity. The reasons are given as follows: First, we We select a two-ray ground model for simulation in NS-2 . By are only concerned with the minimum node lifetime or the connection using the same method, d1 can also be calculated. For the three lifetime in a route from (2). Since two nodes of a stable connection are unknown parameters (t, v, θ) in Fig. 2, we ﬁrst state them in the within the communication range of each other, the connection lifetime following when using the law of cosines to solve the triangle problem may last longer, and they are not a bottleneck from the route to which in Fig. 2: they belong. Second, it is easier to model the mobility of nodes in a short period during which unstable connections last. We can assume d2 = d2 + [v(t1 − t0 )]2 − 2d0 v(t1 − t0 ) cos θ 1 0 (6) reasonably and simply that the nodes move at a constant speed toward the same direction in such a short period. Note that previous works R1 = d2 + [v(t − t0 )]2 − 2d0 v(t − t0 ) cos θ 2 0 (7) – are based on this assumption. The connection time TCi depends on the relative motion between where v = |vi |, and R is the radio-transmission range and is a constant Ni and Ni−1 , and the connection is said to be broken when two that is determined by the system. Another important equation can be nodes (Ni−1 , Ni ) are moving out of each other’s radio transmission derived by the area relation among three triangles ( OAB, OBC, range R. Apparently, there are two important issues here. One is and OAC). In Fig. 2, there is SOAC = SOAB + SOBC , and it is how to measure the distance between nodes Ni and Ni−1 , while the converted into the following when we apply Heron’s formula to it other is how to compute the relative velocity of these two mobile [s = l(l − a)(l − b)(l − c), where the term s represents the area of nodes. a triangle, the terms a, b, and c represent the three sides of the triangle, It is easy to measure the distance between nodes Ni and Ni−1 and l = (a + b + c)/2]: when we use Global-Positioning-System-based location information and then compute it just as described in . Another simple method, l0 (l0 − d0 )(l0 − R) (l0 − v(t − t0 )) which is our approach, is to measure the received signal strength. Assuming that senders transmit packets with the same power level, = l1 (l1 − d0 )(l1 − d1 ) (l1 − v(t1 − t0 )) a receiver can measure the received signal power strength when receiving a packet and then calculates the distance by directly applying the radio propagation model . If the received signal power strength + l2 (l2 − d1 )(l2 − R) (l2 − v(t − t1 )) (8) is lower than a threshold value, we regard this link as an unstable state and then calculate the connection time. where l0 = (d0 + R + v(t − t0 ))/2, l1 = (d0 + d1 + v(t1 − t0 ))/2, Fig. 1 shows the relative motion of two nodes (Ni−1 , Ni ) moving at and l2 = (d1 + R + v(t − t1 ))/2 are all formulated by v and t, and relative velocities vi and vi−1 relative to ground at a given time t. The then, there are three unknown parameters (t, v, θ) in (6)–(8). Thus, ground is used as a reference frame by default. If we consider node Ni the connection breakage time t can be obtained by solving these as the reference frame, node Ni−1 is moving at a relative velocity of v, simultaneous equations, and the residual connection time TCi is as given by the following: calculated as v = vi−1 − vi . (5) TCi = t − t1 . (9)Authorized licensed use limited to: University of Science and Technology of China. Downloaded on March 18,2010 at 04:40:01 EDT from IEEE Xplore. Restrictions apply.
1570 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 59, NO. 3, MARCH 2010 III. D ESCRIPTION OF THE P ROPOSED E XPLORING DYNAMIC NATURE ROUTING P ROTOCOL Since DSR  is one of the most popular routing protocols in MANETs and it is easy to extend the routing control message format of DSR, we implement the proposed route lifetime-prediction algorithm in the DSR protocol. The proposed algorithm consists of the following three phases: route discovery, data forwarding, and route maintenance. There are three main differences between the EDNR and the DSR. First, in the EDNR protocol, every node saves the received signal strength and the received time of the RREQ packet in its local memory, and adds this information into the RREP packet header in a piggyback manner when it receives the RREP for the corresponding RREQ packet to meet the requirement of the connection lifetime-prediction algorithm. Second, node agents need to update their predicted node lifetime during every period. Finally, the node-lifetime information in the RREP packet is updated when the RREP packet is returned from a destination node to the source node. At every EDNR node agent, a variable NLT, which represents the node lifetime, is added to represent the estimated lifetime of this node, Fig. 3. Throughput. and it is updated by the algorithm in Section II-A. For the lifetime of a link Ci , there are two sample packets exchanged between nodes Ni−1 RREQ RREP and Ni (packet 1: Ni−1 − − − −→ Ni ; packet 2: Ni−1 ←− − Ni ) in −− the route-discovery phase, and thus, we can estimate the LLT using the proposed algorithm presented in Section II-B. To implement this, every node agent needs to maintain a data structure called RREQ_Info table in its local memory. This structure includes the RREQ id, the forwarding RREQ time, and the RREQ received signal strength. For a path sequence S, . . . , Ni−1 , Ni , Ni+1 , . . . , D, when an intermediate node Ni receives an RREQ packet from Ni−1 , it adds this RREQ id, the current time, and the received signal strength to its RREQ_Info table before it continues to forward this RREQ packet. Similarly, node Ni+1 also saves the RREQ_Info from node Ni in its local memory. In the returning RREP period, when node Ni receives an RREP packet from node Ni+1 , the RREQ_Info from Ni (information of RREQ Ni − − − −→ Ni+1 ) has been added to the RREP header by Ni+1 before node Ni+1 sends an RREP packet to node Ni . Simultaneously, node Ni knows the RREP time and the RREP received signal strength from RREP node Ni+1 (information of Ni ←− − Ni+1 ). Thus, it can obtain the −− Fig. 4. Number of routing failures. second sample packet that is delivered between the corresponding two nodes (Ni , Ni+1 ), and, thus, we can calculate the connection time TCi using the connection lifetime-prediction algorithm and then update the local LLT value. Similarly, node Ni should add the RREQ_Info entry IV. P ERFORMANCE E VALUATION that is received from node Ni−1 to the RREP header before sending the RREP to node Ni−1 , and then node Ni−1 calculates the LLT between A. Simulation Environment nodes (Ni−1 , Ni ). For our simulations, we use a discrete event-driven simulator NS-2 Three new entries, i.e., path lifetime (PLT), RREQ time, and RREQ . To evaluate the effect of mobility on the performance of routing signal strength, are added to the common header of an RREP packet. protocols, the mobility of nodes follows a random way-point model The PLT represents the predicted lifetime of the source route in this . A mobile node starts a trip to a random destination at a con- packet header and can be updated when RREP packets are forwarded stant speed chosen from a uniform distribution (1, max speed] , from the destination node to the source node in the route-discovery then stops for a predeﬁned pause time, and starts another trip to a phase. The RREQ time and the RREQ signal strength represent the random destination again. The source–destination connection patterns RREQ_Info of the previous RREQ node. The EDNR node agent only are generated using cbrgen.tcl in NS-2. The initial energy is the same updates the PLT value in the common header of the RREP packet with as , and simulation time is set to 1000 s, which is different from a local NLT value or LLT value, if NLT < PLT or LLT < PLT, before that in  because we employ the time of the ﬁrst node breakdown forwarding this RREP packet. When this RREP packet reaches the as the criterion of the network lifetime, whereas the network lifetime source node, the PLT becomes the minimum value of the estimated in  is the time taken for a ﬁxed percentage of the nodes to die. lifetime of all nodes and links through the route from the source node Each data point represents an average of ﬁve different randomly to the destination node, as described in (2). In the persistent data- generated mobility models. The conﬁdence level is set to 95%, and the forwarding period, a source node tends to select the path with the conﬁdence interval is shown as a vertical bar in the ﬁgures (Figs. 3–5). longest lifetime (the path with the maximum PLT value) from multiple Table I summarizes the more detailed simulation parameter paths as a source route for data forwarding. values.Authorized licensed use limited to: University of Science and Technology of China. Downloaded on March 18,2010 at 04:40:01 EDT from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 59, NO. 3, MARCH 2010 1571 Fig. 3 shows the throughput performance in terms of the number of packets for the four routing protocols. The proposed EDNR protocol outperforms the remaining three protocols in varying node velocity en- vironments. Its throughput enhancement is achieved by approximately 79.2%, 14.2%, and 13.8%, compared with that of the original DSR, LPR, and SSA mechanisms, respectively. Fig. 4 shows the advantage of the EDNR protocol in terms of the number of routing failures. To adapt to dynamically varying network topology environments, the EDNR, LPR, and SSA protocols do their best to ﬁnd a more stable route, reducing the number of routing failures by 21.2%, 15.6%, and 14.2%, respectively, compared with that of the original DSR. Fig. 4 shows that the number of routing failures greatly increases with an increase in the node’s mobility speed. Routing overhead is deﬁned as the amount of routing control packets, including RREQ and RREP. Fig. 5 shows the routing overhead of the four routing protocols. The EDNR protocol yields a signiﬁcant improvement with the help of our proposed route lifetime-prediction Fig. 5. Routing overhead. algorithm, and its overhead is reduced by 25.6%, 9.4%, and 6.3%, compared with that of the original DSR, LPR, and SSA, respectively. However, the length of RREP packets is 3 × 4 B longer than that of TABLE I S IMULATION PARAMETERS the DSR. V. C ONCLUSION In MANETs, a link is formed by two adjacent mobile nodes, which have limited battery energy and can roam freely, and the link is said to be broken if any of the nodes dies because they run out of energy or they move out of each other’s communication range. In this paper, we have considered both the node lifetime and the LLT to predict the route lifetime and have proposed a new algorithm that explores the dynamic nature of mobile nodes, such as the energy drain rate and the relative motion estimation rate of nodes, to evaluate the node lifetime and the LLT. Combining these two metrics by using our proposed route lifetime-prediction algorithm, we can select the least dynamic route with the longest lifetime for persistent data forwarding. Finally, we have evaluated the performance of the proposed EDNR protocol based on the DSR. Simulation results show that the EDNR protocol outperforms the DSR protocol implemented with LPR and SSA mechanisms. R EFERENCES  X. H. Wei, G. L. Chen, Y. Y. Wan, and X. M. Zhang, “Longest lifetime path in mobile ad hoc networks,” J. Softw., vol. 17, no. 3, pp. 498–508, 2006. B. Simulation Results  N. Shrestha and B. Mans, “Exploiting overhearing: Flow-aware routing for improved lifetime in ad hoc networks,” in Proc. IEEE Int. Conf. To evaluate the performance of the EDNR, we compare the per- Mobile Ad-hoc Sens. Syst., 2007, pp. 1–5. formance of the EDNR with those of the following three routing  V. Marbukh and M. Subbarao, “Framework for maximum survivability protocols: 1) the original DSR, 2) the DSR with the existing LPR routing for a MANET,” in Proc. MILCOM, 2000, pp. 282–286.  C.-K. Toh, “Maximum battery life routing to support ubiquitous mobile mechanism , and 3) the DSR with SSA , in terms of network computing in wireless ad hoc networks,” IEEE Commun. Mag., vol. 39, throughput, routing failures, and control packet overhead. The original no. 6, pp. 138–147, Jun. 2001. DSR tends to ﬁnd the shortest path from the source node to the  A. Misra and S. Banerjee, “MRPC: Maximizing network lifetime for destination node, ignoring the node lifetime and wireless LLT. The reliable routing in wireless environments,” in Proc. IEEE WCNC, 2002, pp. 800–806. LPR mechanism attempts to ﬁnd a stable route by considering the  M. Maleki, K. Dantu, and M. Pedram, “Lifetime prediction node lifetime based on residual battery and past activity. The SSA routing in mobile ad hoc networks,” in Proc. IEEE WCNC, 2003, mechanism estimates the wireless LLT by classifying neighbors into pp. 1185–1190. a strongly connected set and a weakly connected set, and it drops  C. K. Toh, “Associativity-based routing for ad hoc mobile networks,” the RREQ from weakly connected neighbors. Generally speaking, the Wirel. Pers. Commun.—Special Issue on Mobile Networking and Com- puting Systems, vol. 4, no. 2, pp. 103–139, Mar. 1997. LPR mechanism performs better in a low-mobility scenario, whereas  R. Dube, C. D. Rais, K. Y. Wang, and S. K. Tipathi, “Signal stability- the SSA mechanism seems to be more suitable for a high-mobility based adaptive routing (SSA) for ad hoc mobile networks,” IEEE Pers. scenario. Commun., vol. 4, no. 1, pp. 36–45, Feb. 1997.Authorized licensed use limited to: University of Science and Technology of China. Downloaded on March 18,2010 at 04:40:01 EDT from IEEE Xplore. Restrictions apply.
1572 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 59, NO. 3, MARCH 2010  O. Tickoo, S. Raghunath, and S. Kalyanaraman, “Route fragility: A novel  X. Wu, H. R. Sadjadpour, and J. J. Garcia-Luna-Aceves, “An analytical metric for route selection in mobile ad hoc networks,” in Proc. IEEE framework for the characterization of link dynamics in MANETs,” in ICON, 2003, pp. 537–542. Proc. IEEE Mil. Commun. Conf., 2006, pp. 1–7.  M. Gerharz, C. de Waal, M. Frank, and P. Martini, “Link stability in  D. Johnson, Y. Hu, and D. Maltz, DSR:RFC 4728. [Online]. Available: mobile wireless ad hoc networks,” in Proc. 27th Annu. IEEE Conf. Local http://www.ietf.org/rfc/rfc4728.txt Comput. Netw., 2002, pp. 30–42.  T. K. Sarkar, Z. Ji, K. Kim, A. Medouri, and M. Salazar-Palma, “A  L. Qin and T. Kunz, “Pro-active route maintenance in DSR,” ACM survey of various propagation models for mobile communication,” IEEE SIGMOBILE Mobile Comput. Commun. Rev., vol. 6, no. 3, pp. 79–89, Antennas Propag. Mag., vol. 45, no. 3, pp. 51–82, Jun. 2003. Jul. 2002.  [Online]. Available: http://www.isi.edu/nsnam/ns  W. Su, S. J. Lee, and M. Gerla, “Mobility prediction and routing in  C. Bettstetter, G. Resta, and P. Santi, “The node distribution of the random ad hoc wireless networks,” Int. J. Netw. Manage., vol. 11, no. 1, pp. 3– waypoint mobility model for wireless ad hoc networks,” IEEE Trans. 30, Jan./Feb. 2001. Mobile Comput., vol. 2, no. 3, pp. 257–269, Jul.–Sep. 2003.  P. Samar and S. B. Wicker, “On the behavior of communication links of  J.-Y. Le Boudec and M. Vojnovic, “Perfect simulation and station- a node in a multi-hop mobile environment,” in Proc. Int. Symp. Mobile arity of a class of mobility models,” in Proc. IEEE INFOCOM, 2005, Ad Hoc Netw. Comput., 2004, pp. 145–156. pp. 2743–2754.Authorized licensed use limited to: University of Science and Technology of China. Downloaded on March 18,2010 at 04:40:01 EDT from IEEE Xplore. Restrictions apply.