Frame ork for Reliable Routing in Mobile Ad Hoc Netw orks Zhenqiang Department of Electrical Engineering Uni ersity of California Ri erside zyecs

Frame ork for Reliable Routing in Mobile Ad Hoc Netw orks Zhenqiang Department of Electrical Engineering Uni ersity of California Ri erside zyecs - Description

ucr edu Srikanth Krishnamurthy Satish K ripathi Department of Computer Science and Engineering Uni ersity of California Ri erside krishtripathicsucr edu Abstract Mobile ad hoc netw orks consist of nodes that ar often vulnerable to failur e As such it ID: 29212 Download Pdf

124K - views

Frame ork for Reliable Routing in Mobile Ad Hoc Netw orks Zhenqiang Department of Electrical Engineering Uni ersity of California Ri erside zyecs

ucr edu Srikanth Krishnamurthy Satish K ripathi Department of Computer Science and Engineering Uni ersity of California Ri erside krishtripathicsucr edu Abstract Mobile ad hoc netw orks consist of nodes that ar often vulnerable to failur e As such it

Similar presentations


Download Pdf

Frame ork for Reliable Routing in Mobile Ad Hoc Netw orks Zhenqiang Department of Electrical Engineering Uni ersity of California Ri erside zyecs




Download Pdf - The PPT/PDF document "Frame ork for Reliable Routing in Mobile..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.



Presentation on theme: "Frame ork for Reliable Routing in Mobile Ad Hoc Netw orks Zhenqiang Department of Electrical Engineering Uni ersity of California Ri erside zyecs"ā€” Presentation transcript:


Page 1
Frame ork for Reliable Routing in Mobile Ad Hoc Netw orks Zhenqiang Department of Electrical Engineering Uni ersity of California, Ri erside zye@cs.ucr .edu Srikanth Krishnamurthy Satish K. ripathi Department of Computer Science and Engineering Uni ersity of California, Ri erside krish,tripathi@cs.ucr .edu Abstract Mobile ad hoc netw orks consist of nodes that ar often vulnerable to failur e. As such, it is important to pr vide edundancy in terms of pr viding multiple node- disjoint paths fr om sour ce to destination. first pr o- pose modified ersion of the popular

OD pr otocol that allo ws us to disco er multiple node-disjoint paths fr om sour ce to destination. find that ery few of such paths can be ound. Furthermor e, as distances between sour ces and destinations incr ease, bottlenecks ine vitably occur and thus, the possibility of finding multiple paths is considerably educed. conclude that it is necessary to place what we call reliable nodes (in terms of both being ob ust to failur and being secur e) in the netw ork or efficient operations. pr opose deployment strategy that determines the po- sitions and the trajectories of these

eliable nodes such that we can achie framew ork or eliably outing inf orma- tion. define notion of reliable path which is made up of multiple segments, each of which either entir ely con- sists of eliable nodes, or contains pr eset number of mul- tiple paths between the end points of the segment. sho that the pr obability of establishing eliable path between random sour ce and destination pair incr eases considerably en with lo per centage of eliable nodes when we con- tr ol their positions and trajectories in accordance with our algorithm. Mobile ad hoc netw orks find application

in man fields such as military deplo yments, disaster rescue missions, electronic classrooms. In this paper we primarily look at reliability in terms of pro viding rob ustness to node ail- ures in ad hoc netw orks. Node ailures may be intermit- tent, i.e., for short periods or for long periods of time, and due to arious reasons. First, since these netw orks are lik ely to be deplo yed in wireless en vironments, the com- munications between the ad hoc nodes will ha to be via harsh ading channel. Thus, communications between This ork as supported by ARP under contract number: FTN

F30602-01-2-0535. nodes ould typically endure periods of intermittent ail- ure and as consequence, pack et losses. It is possible that certain nodes might completely lose connecti vity for temporary periods due to the ading conditions. One ay of ercoming this ould be to use sophisticated antenna systems or modulation methods. Ho we er man of the ad hoc nodes, if not most of them, ould be constrained by size, processing and po wer limitations and thus, may not possess such capabilities. Second, man of the ad hoc nodes are po wer constrained. Due to battery drain, it is possible that some of

these nodes might not be able to function. Such an ef fect may result in long term ailure if node battery is completely drained or if it is possible to re-char ge the node battery the node might not func- tion for intermittent short periods. Third, nodes in an ad hoc netw ork are vulnerable to compromise. Compromises are especially lik ely for unattended sensor nodes or hand- helds carried by pedestrians. simple form of denial of service is to simply cause node ailures, either intermittent or long term. Multipath routing is one ay of impro ving the relia- bility of the transmitted information.

While multipath routing may be used for arious other reasons such as load-balancing, congestion oidance, lo wer frequenc of route inquiries and to achie lo wer erall routing er head [1][2][3][4][5], our objecti is to primarily design multipath routing frame ork for pro viding enhanced ro- ustness to node ailures. If one could pro vide multiple paths from source to destination, one could en vision the transmission of redundant information on the arious paths (by the use of kno wn techniques such as di ersity encoding [6]) that ould help the recei er in reconstruct- ing the transmitted

information en if fe of the paths were to ail. By multiple paths, we imply multiple node- disjoint routes from source node to destination node. Our first goal to ards this is to design routing protocol that ould allo us to find multiple node-disjoint paths from gi en source to destination. ards this, we
Page 2
mak modifications to the Ad Hoc Distance ector Rout- ing Protocol (A OD V) [7] which is one of the most popu- lar ad hoc routing protocols to acilitate the disco ery and consequently the use of multiple node-disjoint paths. found that the number of

node-disjoint paths from source to destination is dependent on the node density in the ad hoc netw ork (as might be xpected). Furthermore, we found that as the distance between source and its des- tination is increased, one could find no more than ery limited number of paths between them, en at moderate node densities (a erage node de gree is 6.7). This observ a- tion lead us to belie that, one ould require at least fe of the ad hoc nodes to be more eliable One could en vi- sion that these nodes ould be placed in mo ving ehicles and could be less constrained in terms of size, processing

and po wer The ould be physically more secure and ro- ust to compromises. These nodes (typically much fe wer in number in comparison with the normal ad hoc nodes) could then, be allo wed to participate in routing along mul- tiple routes between the same source-destination pair or the ease of notation let us call these nodes R-nodes The re vised objecti is then to construct sequence of e- liable se gments between the source and the destination. Nodes that join tw se gments ha to be R-nodes. se g- ment is deemed reliable if it consists of either preset number of paths between the tw R-nodes that

it connects or if it is made up of R-nodes entirely concatenation of reliable se gments is called eliable path describe the construction of reliable path in detail in Section V. The ne xt question that arises is: where should these R- nodes be placed so that the probability of finding reli- able path between an arbitrary source and destination is acceptable? Initially we placed these R-nodes at random locations within the area of interest. Ho we er we found that this does not help in achie ving an acceptable prob- ability of finding reliable path between source and destination.

Thus, we need more intelligent ay of plac- ing these R-nodes. Furthermore, as the nodes in the ad hoc netw ork are mobile, the R-nodes ould ha to adapti ely mo so as to maintain these adv antageous positions with respect to the other nodes. propose methodology to control the trajectory of an R-node based on information xchanged within local vicinity of the R-node. find by simulations that placing each R-node at positions de- fined by our algorithm (which is in act, ersion of the randomized min-cut algorithm [8]) is ery ef fecti de- plo yment strate gy in terms of achie ving high

probability that reliable path is found between an arbitrary source and destination. The remainder of this paper is or ganized as follo ws. In The details are pro vided in later section. Section II, we re vie the related ork on multipath rout- ing in ad hoc netw orks. describe our modified er sion of OD (we call it OD VM for OD -Multipath) in Section III and describe ho it finds multiple node- disjoint paths from gi en source to gi en destination. In Section IV we discuss the simulation xperiments per formed with OD VM and discuss the observ ed results in terms of performance.

describe the arious strate gies that we consider for deplo ying the R-nodes and the mo- ti ation for doing so in Section V. In Section VI we de- scribe our simulation results with ne xperimental set up with an ad hoc netw ork that includes small number of R-nodes and discuss the observ ed results in terms of the performance of the arious deplo yment strate gies. present our conclusions in Section VII. Multipath routing has been well studied in wired [1][9][10] and wireless [2][3][4][5][11] netw orks. Mul- tipath routing in MANETs has also recei ed some atten- tion recently DSR [12] and ORA

[13] ha the abil- ity to find multiple paths. In DSR, by using the infor mation recei ed from multiple route queries which might tra erse distinct paths, the destination can attempt to con- struct multiple node-disjoint paths. Ho we er due to its inherent nature (as in OD described in the ne xt sec- tion), DSR can find only small fraction of the possi- ble node-disjoint paths if used without an modifications. ORA uilds and maintains multiple loop free paths using Directed Ac yclic Graph (D G) rooted at the destination; ho we er it does not find node-disjoint paths. ath

disjointness has been studied in [2][3][5][11]. In [2], the authors ha analyzed the performance impacts of alternati path routing for load balancing. Nasipuri et al.[3] studied the ef fect of number of multiple paths and lengths of those paths on routing performance using an- alytical models. Lee et al. [11] proposed the Split Mul- tipath Routing protocol (SMR), which can find an alter nate route that is maximally disjoint from the shortest de- lay route from the source to the destination. All of the abo protocols are based on source routing. Distance ector based multipath routing

protocols are in estigated in [4][9][14]. Ho we er of these, OMD [4] is the only protocol that ensures that the paths are edge-disjoint. The multipath routing protocols described abo e, which are based on source routing, allo the source node to com- pute multiple node or edge-disjoint paths. The source can do so from the partial topology information that is made ailable by means of multiple responses to single route query ith distance ector based protocols, the topol- ogy information that node can obtain is further limited.
Page 3
Thus, it ould be dif ficult to construct

node-disjoint paths from source to destination. Link state routing can be used to generate multiple node-disjoint paths ut the use of such protocols requires lar ge erheads [15]. OD is popular routing protocol that creates distance ector routing tables on-demand and it requires lo wer erhead as compared with DSR [16]. Thus, we choose OD as candidate protocol and mak modifications to it, to a- cilitate the disco ery of node-disjoint paths from source to destination. Although there has been prior ork on modifying OD to compute edge-disjoint paths [4], to the best of our kno wledge, our OD

VM protocol is the first modified ersion of OD that has the ability of finding node-disjoint paths. Furthermore, our ork is the first to study the relationship between the number of node- disjoint paths that can be found between source and destination and the density of nodes in the netw ork. Our observ ations lead us to conclude that in the absence of an infrastructure it is highly improbable that we can find satisf actory number of node-disjoint paths en at mod- erate densities, especially when the source and the desti- nation are ar apart. Thus, we propose an

infrastructure that is acilitated the deplo yment of reliable nodes (that we call R-nodes ), that can route on multiple paths, as de- scribed earlier Our ork also in estigates the ef fect of the location of R-nodes on the performance in terms of computing mul- tiple paths. propose distrib uted protocol to control the trajectories of the R-nodes such that reliable rout- ing frame ork could be pro vided. In [17], trajectory control algorithm as proposed for mobile gate ays in ad hoc netw orks. The objecti of the trajectory control algo- rithm is to determine where the gate ays are to be placed,

relati to the ad hoc group of nodes that the gate ay serv es such that certain netw ork performance metrics such as throughput as maximized. Unlik in [17] wherein one ould most lik ely place the gate ays in dense re gions within the netw ork, our objecti ould be to place the R- nodes in sparser re gions of the netw ork and control their trajectories so as to increase the probability of establish- ing reliable path (defined earlier) between tw arbitrary nodes. DV In order to acilitate the computation of multiple node- disjoint paths from source to the destination, we choose OD as

candidate protocol and mak modifications to it to enable the disco ery of such paths. First, the choice of OD is based on prior studies [15] that sho that on- demand routing protocols consume lo wer erhead than Source Destination Links that are discarded Links that are recorded in the routing table Fig. 1. The RREQ propagation procedure in OD pro-acti routing protocols. Second, as compared with DSR (which is the other popular on-demand routing pro- tocol), OD oids the high source routing erhead. A. OD first briefly describe the OD protocol. omit most of the details due to

space limitations. more de- tailed description of OD may be found in [7]. OD combines the use of destination sequence num- bers in DSD with an on-demand route disco ery tech- nique. If source needs route to destination, it in ok es netw ork-wide flood of route request or RREQ mes- sage. In response, either the destination or an intermediate node that kno ws route to the destination, sends route reply or RREP message back to the source along the path on which the RREQ message as recei ed. Intermediate nodes re-broadcast the RREQ message only if (a) the do not kno route to the destination

and (b) if the ha not already forw arded the particular RREQ message. Once route is established, it is used by the source to send data. If link ails, the node that detects the link ailure (possibly through feedback from the link layer), sends route error (RERR) message to the source, upon the receipt of which, the source re-initiates route search. Destination sequence numbers are tagged onto all routing messages and are used to indicate the relati fr eshness of the routing information. Since duplicate RREQ messages are discarded by in- termediate nodes, it is probable that, some of the possi-

ble node-disjoint paths to the destination, might ne er be traced during the query process. In Fig. 1, the links indi- cated by the dashed lines are ne er reported to the destina- tion since the intermediate relay nodes discard the RREQ messages recei ed on these links. Ev en though there are three possible node-disjoint paths from the source to the destination, OD can find only one of them. B. OD -Multipath (A OD VM) propose modifications to the OD protocol so as to enable the disco ery of multiple node-disjoint paths from source to destination. Instead of discarding the duplicate

RREQ pack ets, intermediate nodes are required
Page 4
Destination ID Source ID Neighbor List Hops To Source Expiration Timer Neighbor ID Destination Sequence Number Destination ID Route List Source Sequence Number Expiration Timer Last Hop ID Source ID Next Hop ID Hop Count (b) (a) Fig. 2. (a) Structure of the each RREQ table entry in OD VM (b) Structure of the each routing table entry in OD VM to record the information contained in these pack ets in table which we refer to as the RREQ table. or each re- cei ed cop of an RREQ message, the recei ving interme- diate node records the

source who generated the RREQ, the destination for which the RREQ is intended, the neigh- bor who transmitted the RREQ, and some additional in- formation (as sho wn in Fig. 2(a)) in the RREQ table. Furthermore, intermediate relay nodes are precluded from sending an RREP message directly to the source. When the destination recei es the first RREQ pack et from one of its neighbors, it updates its sequence num- ber and generates an RREP pack et. The RREP pack et contains an additional field called last hop ID to indi- cate the neighbor from which the particular cop of RREQ pack et as

recei ed. This RREP pack et is sent back to the source via the path tra ersed by the RREQ cop albeit in the re erse direction. When the destination recei es dupli- cate copies of the RREQ pack et from other neighbors, it updates its sequence number and generates RREP pack ets for each of them. Lik the first RREP pack et, these RREP pack ets also contain their respecti last hop nodes IDs. When an intermediate node recei es an RREP pack et from one of its neighbors, it deletes the entry correspond- ing to this neighbor from its RREQ table and adds rout- ing entry to its routing table (sho

wn in Fig. 2(b)) to indi- cate the disco ered route to the originator of the RREP pack et (the destination). The node, then, identifies the neighbor in the RREQ table via which, the path to the source is the shortest, and forw ards the RREP message to that neighbor The entry corresponding to this neighbor is then deleted from the RREQ table. In order to ensure that node does not participate in multiple paths, when nodes erhear an node broadcasting an RREP message, the delete the entry corresponding to the transmitting node from their RREQ tables. When an intermediate node that recei es

an RREP mes- sage cannot forw ard it further (its RREQ table is no empty), it generates an RDER or Route Disco ery Error message and sends that message to the neighbor that actu- ally forw arded the RREP to this node. The neighbor upon assume that the ID of node is unique in the netw ork and it can be the node IP address. recei ving the RDER message will no attempt to forw ard the RREP to dif ferent neighbor who can potentially for ard it further to ards the source. limit the number of RDERs that an RREP message can xperience in order to pre ent the generation and xchange of lar ge number of

such pack ets see that intermediate nodes mak decisions on where to forw ard the RREP messages (unlik in source routing) and the destination, which is in act the origi- nator of these messages is una are as to ho man of these RREP messages that it generated actually made it back to the source. Thus, it is necessary for the source to confirm each recei ed RREP message by means of Route Confirmation message (RRCM). The RRCM mes- sage can, in act, be piggyback ed onto the first data pack et sent on the corresponding route and will also contain in- formation with re gards to the

hop count of the route, and the first and last hop relays on that route. As in the OD protocol, we use sequence numbers to pre ent loops. When source node initiates an RREQ, it increases its sequence number    represents node latest sequence number kno wn to node and the destination sequence number     by one. These tw sequence numbers are indicated in the RREQ pack et and denoted by    and     respecti ely Each time the destination node recei es an RREQ pack et, it computes ne sequence number:       !#"%$'&   

)(      *,+.- (1) The destination then generates an RREP message that contains sequence number    0/ which is set to      Lemma Using OD VM, if oute 13254 ..., ,..., 2687 is found, wher is the :9 node on the path, is the sour ce node (the originator of the RREQ query) and 26 is the destination, then <; for any and ->= =@? i.e ., ther is no loop in this oute Pr oof When node forw ards an RREP pack et to ards the source, it adds an entry in its routing table to indicate route from the destination to the source. Assume that there is loop on the route,

ithout loss of generality we assume that is on the loop. Thus, ould forw ard the same RREP message more than once. ith OD VM, when node forw ards an RREP it “implicitly informs all its neighbors that it is part of the corresponding route. Upon the receipt of this message, the node neighbors delete the entry corresponding to the transmitting node in their RREQ tables. Thus, when transmitted the RREP message to BA all of the neighbors of that erheard In our simulation (to be described later) we set this limit to twice the lifetime (TTL) of the RREP pack et.
Page 5
the RREP ould delete

the entry corresponding to in their RREQ tables. Thus, these nodes ould ne er for ard another RREP to If node ailed to erhear RREP message, it is possible that it may forw ard an RREP to Ho we er upon the receipt of this RREP since is already on an acti route, it cannot forw ard the RREP to an other neighbor; ould send an RDER message to the particular neighbor Thus, the loop is pre ented. Lemma Using OD VM, if two outes 1D2  ,..., ,...,   and 1@2  ,..., 2E ,...,   ar disco ver ed, these two outes have no common nodes xcept for the sour ce  and the destination   i.e ., these two

outes do not contain any common intermediate nodes and ar hence node-disjoint. Pr oof Since, from Lemma 1, node ne er forw ards more than one RREP in response to the same RREQ, it is impossible for node to participate in more than one route. Thus, if multiple routes are disco ered, the should be node-disjoint. One of the disadv antages with OD VM is that interme- diate nodes cannot use pre viously cached routing infor mation to generate RREP messages. The RREP messages should al ays be generated by the destination node. This, ho we er is necessary since, if intermediate nodes gener ate RREPs,

it might not be feasible to guarantee that the disco ered routes are node-disjoint. DV In this section, we aluate the performance of the OD VM protocol and discuss the ailability of multi- ple node-disjoint paths with arious node densities. use simulation model based on ns-2 [18]. The Monarch research group in CMU de eloped support for simulating multi-hop wireless netw orks complete with physical, data link and MA layer models in ns-2 The distrib uted co- ordination function (DCF) of IEEE 802.11 for wireless LANs is used as the MA layer The radio model uses characteristics similar to

commercial radio interf ace, Lu- cent eLAN. eLAN is shared-media radio with nominal bit-rate of 2Mb/sec and nominal radio range of 250 meters. The performance metrics that we are inter ested in are: The erage number of node-disjoint paths that are disco ered per route inquiry The probability that the number of node-disjoint paths disco ered in an route inquiry is no less than certain preset threshold In our simulations we disperse arying number of nodes (Case 1: 250 nodes, Case 350 nodes and Case 3: 500 nodes) uniformly in 2500m 2500m rectangular re- gion. use the random aypoint model to model

node mo ements. ause time is al ays set to zero and the speed of the nodes is uniformly distrib uted er IJLKM In each case, we generate 20 dif ferent scenarios. In each scenario, we randomly choose 500 source and destina- tion pairs. The simulation results are the erage of these 10000 samples. Since ery RREP pack et tries to find the shortest path from the destination to the source, note that the number of node-disjoint paths disco ered by OD VM is not the maximal number of node-disjoint paths that can be found between the source and the destination. Ho we er without xpending lar ge

amount of erhead in order to obtain the topology information of the entire netw ork, it is im- possible to compute all the node-disjoint paths. In order to aluate the performance of OD VM, we compare it with an ideal case, in which the topology of the entire net- ork is kno wn at the source and the source first ecutes the shortest path first search algorithm. The nodes on the shortest path are no xcluded and the algorithm is e- cuted again to compute the ne xt shortest path. Note that this ne path is node-disjoint from the first path. The pro- cess is then repeated until no

further node-disjoint paths can be found between the gi en source and destination. In Fig. the performance of OD VM is compared with that of the ideal case while arying the density of nodes in the netw ork. In Cases and Case 3, OD VM can find at least 80% of the paths found in the ideal case, and it can find at least 70% of the paths found by the ideal method in Case 1. The higher the node density the higher this percentage. This is because the higher the node den- sity the higher the probability that multiple paths xist between the source and the destination. At lo wer node

densities, there may xist some “bottleneck nodes in re- gions of lo node density between the source and the des- tination. Since these nodes can only route pack ets on single path, other RREPs ha to mak detours and find alternate routes. Ho we er we note that there is limit im- posed on the number of RDERs that an RREP pack et can xperience. Furthermore, some of the RREQ messages are lost due to collisions and hence, do not result in RREP responses. Due to these ef fects, some alternate paths (e en if the xist) may ne er be found. Fig. (and Fig. 5) plot the probability that the num- ber

of node-disjoint paths disco ered in each route inquiry by OD VM is no less than (and 4) ersus the number of hops on the shortest path between the source and the destination. From Fig. 4, we see that the probability that at least three paths are found is almost in Case 3, and is abo 0.78 in Case 2. But in Case 1, this probability drops quickly as the distance between tw nodes increases. In measure of the distance between the tw nodes.
Page 6
10 10 12 14 Number of hops of the shortest path between two nodes Average number of node−disjoint paths Case 1 (250 nodes): AODVM Case 1

(250 nodes): Ideal search Case 2 (350 nodes): AODVM Case 2 (350 nodes): Ideal search Case 3 (500 nodes): AODVM Case 3 (500 nodes): Ideal search Fig. 3. erage number of node-disjoint paths disco ered per route inquiry for arious node densities. 10 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Number of hops of the shortest path between two nodes Probability Case 1 (250 nodes) Case 2 (350 nodes) Case 3 (500 nodes) Fig. 4. Probability that the number of node-disjoint paths disco ered is no less than OQPSR per route inquiry for arious node densities. Fig. 5, the probability that at least four paths are found is abo

0.77 in Case 3. In Case 2, this probability drops quickly to 0.5 as the distance between tw nodes increases to about hops. It drops belo 10% in Case as the dis- tance between the source and the destination is hops. From Fig. 3,Fig. and Fig. we note that when the node density is high, we can find an acceptable number of node-disjoint paths to pro vide reasonable le el of ro- ustness to node ailures. Ho we er the number of node- disjoint paths that are disco ered is ery limited en at moderate node densities (for xample, Case 1). In order to route information reliably in cases wherein,

multiple node-disjoint paths are not ailable, certain number of “reliable nodes should be placed in the netw ork. In the ne xt section we describe the functionality of these reliable nodes and describe methodology to control their trajec- tories to achie higher routing reliability 10 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Number of hops of the shortest path between two nodes Probability Case 1 (250 nodes) Case 2 (350 nodes) Case 3 (500 nodes) Fig. 5. Probability that the number of node-disjoint paths disco ered is no less than OTP8U per route inquiry for arious node densities. In the pre vious

section we sa that, without xpending lar ge amount of erhead, one cannot find suf ficient number of node-disjoint paths between gi en source and destination to pro vide reasonable de gree of rob ustness to node ailures. This as especially true if the source and the destination were ar ay from each other One could immediately think of finding edge-disjoint paths; ho w- er nodes that are at the intersection of multiple routes might ail and this might cause all the routes which pass through such node to ail simultaneously upon the node ailure. Thus, it is concei able that one

ould attempt to deplo those nodes that are more reliable than others at junctions connecting multiple node-disjoint se gments (a se gment is path between tw nodes, see Fig. 6). In this ork, we propose that set of these reliable nodes be deplo yed in an ad hoc netw ork for the purposes of increasing reliability and security This proposition is not unrealistic in the sense that in typical ad hoc deplo y- ments one can en vision the presence of multiple types of nodes. In battlefield netw ork, one could ha unreliable lo po wer sensors or handhelds, whereas there could be the more reliable,

po wer capable and secure nodes that are located in tank or an other lar ge ehicles. It is also com- mon in security research to assume the presence of the so called “trusted nodes [19]. or the ease of discussion, we refer to these reliable nodes as R-nodes It ould be natu- rally xpensi to deplo lar ge number of these R-nodes and the R-nodes ould constitute small fraction of the entire ad hoc netw ork. The question that we are trying to answer is: if the objecti of deplo ying these R-nodes is primarily to support reliable routing frame ork, then, where should these R-nodes be positioned and ho

should their trajectories be controlled? Before we try to answer this question we first define
Page 7
what we call eliable path In the absence of the R- nodes, we state that reliable path xists from source to destination if the number of node-disjoint paths that can be found between this source and destination is at least equal in number to preset threshold When the R-nodes are deplo yed, the definition of eliable path changes. If one can concatenate sequence of eliable se gments between the source and the destination node, then the path is said to be reliable. se

gment is defined to be reliable if the num- ber of node-disjoint paths that can be found between the end nodes of the se gment is at least equal in number to or if the se gment entirety consists of reliable nodes. Note that, while concatenating reliable se gments, the nodes at the intersection of such se gments ought to be reliable As an xample, we ha path from source to destina- tion in Fig. 6. The alue of is set to three. There are three R-nodes XYE and XYZ see that the end-to- end path from to may be deemed reliable since we can concatenate three reliable se gments, the first

from to X[4 the second from X\4 to and finally the last from XYZ to It is important to position the R-Nodes so as to maxi- mize their utility As the nodes in the mobile ad hoc net- ork mo e, it may become necessary to mo the R-nodes relati to the motion of the other nodes. If the netw ork is dense all er it ould be possible (as the results indi- cated with OD VM) to find reasonable number of paths between an arbitrary source and destination pair As an xample, when the erage node de gree as set to 13.5 (Case in Section IV), OD VM as able to find eight paths, on erage. When we

considered lar ge sample of source-destination pairs that are separated by fix ed hop count on the shortest path between them, we found that the minimum and the maximum of the number of paths that are found between all such pairs are significantly dif fer ent from one another Thus, the reason why we could not find multiple paths between nodes that are distant from each other is most probably because of the presence of sparse re gions in the netw ork which act as bottlenec ks If we could place the reliable nodes in these sparse areas, it appears as if we could create the

desired reliable paths. Randomly placing these R-nodes is not lik ely to pro vide us with an performance gains (as we shall see later). By placing these R-nodes such that the ould interconnect with the maximum number of ad hoc nodes (i.e., ha maximal de gree) ould probably not help either as we see in the xample in Fig. 7. In Fig. 7(a), the black node has the lar gest de gree in the netw ork. In Fig. 7(b), the black node has the smallest de gree in the netw ork. Ho we er assume the communicating entities are reliable and ha mutu- ally authenticated themselv es. segment 1 segment 2 segment 3

R−nodes Normal Nodes Fig. 6. reliable path between node and node OQP8R ). (a) (b) Fig. 7. (a) The maximum-de gree node (the black node) is the bottle- neck node in the netw ork. (b) The minimum-de gree node (the black node) is the bottleneck node in the netw ork. each node has an equal importance in terms of eeping the netw ork connected, i.e., ensuring that single path x- ists between an tw nodes. Our objecti is similar i.e., identify positions for the R-nodes such that the probabil- ity of the xistence of eliable path (gi en alue of between an tw gi en nodes is high. ards this, we use

modification of the randomized min-cut algorithm which we describe in the ne xt sub-section. A. Min-cut algorithm and our modification Prior to describing ho the randomized min-cut algo- rithm may be used to determine where the R-nodes are to be placed, we describe the randomized min-cut algorithm [8] in brief. Let & k (ml be an undirected weighted graph which is connected. cut in is partition of the ertices into tw non-empty sets and The alue of cut is the sum of the weights of the edges crossing the cut. If the weights of all the edges in are one, then the alue of cut is the

count of those edges that ha one end-point in each of the tw sets and The min-cut is the cut(s) with the minimum cut alue of all the possible cuts. If all the edges in are of unit weight, the min-cut is the number of edges that must be remo ed from to separate it into tw partitions. The smaller one of these tw partitions is then called min-cut set. cut of gi en graph can be obtained by what is called the contr action algorithm The basic idea of the contr ac- tion algorithm is to randomly choose an edge &on ( p in and replace ertices and by ne erte for each
Page 8
2rK s t ( pu the

weight of the ne edge is the sum of the weights of edge and edge ( p the rest of the graph remains unchanged. The contraction procedure is repeated until there are only tw nodes and one edge left. The cut alue is then the weight of the edge that connects these tw nodes. During each iteration of the contraction procedure, single edge is chosen and the tw nodes con- nected by this edge are contracted. Thus, if there are nodes in the graph, the algorithm tak es time. Note that this running time is independent of the number of edges in the graph. It can be pro ed that the probabil- ity that the

min-cut of graph is found by single run of the contraction algorithm is bounded by [8]. If we repeat the contraction algorithm Eyx{z| ?* times, we can xpect with reasonable probability that some itera- tions of the contraction algorithm find the min-cut. Thus, in order to compute the min-cut alue of gi en graph we ould xpect to incur run time of ?,} x{z| ?* In order to determine where the R-nodes ought to be placed, we require each node to compute the min-cut of partial graph. The objecti is to determine ho vul- nerable the netw ork is, in terms of becoming partitioned if particular

node as remo ed from the graph (i.e., as in ailure). assume that each node can obtain par tial topolo gy vie of the netw ork; more specifically we assume that it kno ws the entire topology within some hops from itself is system design parameter). The node then remo es itself and the edges incident on itself from the graph representing this partial topology It then runs the min-cut algorithm with the follo wing modifica- tion: The outermost links are contracted first, and the links that are closest to the node are contracted last. This is done in an attempt to ensure that the

min-cut is an accu- rate indicator of the importance of the computing node in eeping the localized topology connected. As an xam- ple, in Fig. 8(a), the black node is the one performing the computations; one ould lik the min-cut to in act “pass through the links associated with the black node (sho wn by the dotted lines) as sho wn. ithout the requirement that the outermost edges be contracted first, the min-cut ould probably pass through one of the outermost edges. In Fig. 8(a), without the requirement, the min-cut ould pass through the link between nodes and and its alue is one. This ho

we er does not reflect on the rela- ti importance (which is of interest) of the black node in eeping the graph connected. Fig. 8(b) through 8(j) illustrate one of iterations of the contraction algorithm which finds cut alue of the graph Clearly this is done to estimate the vulnerability of the localized neighborhood in terms of becoming partitioned if the node performing the computation were to ail. 10 11 10 {V , V } 11 11 {V , V } 17 10 (c) (b) (d) (e) (g) 14 14 (h) (f) 11 13 (a) 16 16 14 (j) (i) min−cut of the black node Links connected with the black node are not counted.

{V 10 , V } 12 11 13 {V 15 , V } 16 {V 11 15 {V , V } 14 {V 14 18 {V , V } 12 , V } 13 , V } 17 Fig. 8. An iteration of the contraction algorithm. sho wn in Fig. 8(a). The initial weight of ery edge is one. In each contraction step, an edge (we choose the dashed edges as sho wn is chosen first, among the out- ermost edges, and the tw nodes connected by this edge are contracted. The inner most edges are chosen in the fi- nal fe steps. Finally (Fig. 8(j)), only one edge and tw nodes are left. The alue of the cut as determined by this iteration is tw o. Since our modification

does not change the number of nodes in the input topology and the only dif ference is in the contraction sequence of the nodes, the computation comple xity remains the same as that of the original min- cut algorithm, i.e., x{z| ?* if there are nodes within hops of the node computing the min-cut. If is small, the comple xity may be xpected to be airly lo In the follo wing sub-sections we describe centralized and distrib uted approach of using the abo method- ology to determine the best positions for placing the R- nodes. Although centralized approach is unrealistic within mobile ad hoc netw

ork setting, it is useful in terms of aluating the goodness of our distrib uted algo- rithm. This choice is arbitrary and is done simply for illustrati purposes. could choose the edges in dif ferent order as well.
Page 9
B. Using centr alized contr oller to determine R-node placement In the centralized strate gy we assume that the topology information of the entire netw ork is kno wn to ery node. Ev ery node min-cut alue and min-cut set are computed priori with respect to graph of its local topology up to hops from it. As described earlier we then place the R-nodes in the positions

occupied by the nodes with the lo west min-cut alues. This centralized strate gy requires static netw ork topology and mobility is not allo wed. The performance in terms of the probability that reliable path is found between an arbitrarily chosen pair of nodes, as achie ed by thus the placing of R-nodes, can be used as benchmark to compare with our distrib uted ersion of the R-node placement algorithm. C. The distrib uted R-node deployment str ate gy In the distrib uted R-node deplo yment strate gy we as- sume that each node in the netw ork has information (by using GPS or other techniques)

that specifies its wn co- ordinates. further suppose that ery node periodically broadcasts HELLO message to its neighbors; informa- tion which specifies the topology of the node -hop local neighborhood is included in this HELLO message. If is small, we can xpect that within some short finite time, each node has the complete information about the topol- ogy of its -hop neighborhood. R-nodes transmit HELLO messages as well. In an R-node HELLO message, there is flag that is used to indicate its motion status: static (if this R-node position has been determined) or dynamic

(if this R-node is still in the process of determining where to mo e). Each normal node can thus construct tw lo- cal topology graphs, the first with the static R-nodes and the second without the static R-nodes. The dynamic R- nodes are not included in either of these tw graphs. node periodically calculates its min-cut alue and the size of the min-cut set based on these tw graphs. Note that the weight of direct link between tw static reliable nodes is set to All the other links ha weight of one. The com- puted min-cut alues and the corresponding min-cut set sizes are piggyback ed onto

the node HELLO message. An R-node compares the min-cut alue and the min-cut set sizes of the nodes in its -hop neighborhood, and it mo es to the proximity of the normal node that has the minimum min-cut alue. If the min-cut alues of tw nodes are the same, the reliable node will mo to the proximity of the node that has lar ger min-cut set. In order to pre ent multiple R-nodes from mo ving to the same location at the same time, before an R-node mo es to the proximity of normal node, it sends out motion re- quest to that normal node. The R-node does not mo un- til it recei es motion

confirmation from the normal node. Some additional constraints can also be incorporated, such as requiring that no tw R-nodes can be too close and limiting the number of R-nodes within the range of par ticular R-node D. Modifications to OD VM In order to allo the incorporation of the R-nodes and to allo these nodes to participate in multiple paths, OD VM has to be further modified. Ho we er the changes are ery simple and lightweight. In each RREP pack et, we include what we call elia- bility flag. When the RREP pack et passes through an in- termediate node, this

flag is set to RELIABLE only if this intermediate node is an R-node and if the original alue of this flag as also RELIABLE Otherwise, this flag is set to NORMAL If an intermediate R-node can not find ne xt hop R-node to forw ard this RREP pack et, it will split the RREP pack et into multiple RREP pack ets equal in count to the number of neighbors specified in its RREQ table. All of these RREP pack ets are mark ed NORMAL and then forw arded to the dif ferent neighbors. In the xample in Fig. 6, let us assume that is set to three. Initially node generates and sends

an RREQ message to node Upon recei ving the RREQ 10 node generates an RREP pack et and attempts to send this pack et back to via and When recei es the RREP message (mark ed RELI- ABLE ), it is unable to forw ard it further to reliable node. Being are, that it has actually recei ed three copies of the original RREQ from three normal ad hoc nodes (by means of its RREQ table), it then mak es three copies of the RREP message recei ed from XYE It then marks these messages NORMAL and forw ards one cop to each of the three neighbors. The three RREP copies, then, find their ay to the source.

Since as three, and three RREP messages were recei ed, the source infers that “reliable path is ailable to the destination E. Ef fects of node mobility The mobile ad hoc netw ork topology changes as nodes mo e. In order to maintain the reliable routing frame ork, the R-nodes will ha to correspondingly mo to re vised locations as the netw ork olv es. If the maximum speed of motion of the R-nodes is the same (or lo wer than) as that of the normal nodes, the will not be able to mo specify this distance to be 50m in our simulations. Ho we er this ould be system parameter that can be

configured. This is system parameter as well. Ho we er in our simulations, we found that if this number is set to 4, we observ the best performance. …„ Notice that only single cop of the RREQ is recei ed by the destination.
Page 10
quickly enough to ne strate gic positions in timely man- ner Thus, requirement ould be that the R-nodes should be able to mo at much aster speeds as compared to its normal ad hoc nodes. This is concei able since as men- tioned earlier these R-nodes are typically po werful and housed in lar ge ehicles as opposed to being sensors or being

carried by pedestrians. OY In our simulations, we focus on Case described in sec- tion IV. In this scenario, 250 nodes are deplo yed in rect- angular area of 2500m 2500m. choose this case to demonstrate the ef fecti eness of our R-node deplo yment strate gy en when the density of nodes in the netw ork is moderate. In all our simulation xperiments we choose (the number of paths that ould deem particular se g- ment, made up of normal ad hoc nodes, reliable) to be either or 4. This number seems to be reasonable for the population size considered and we ant to oid x- tremely long paths that are

dif ficult to maintain 11 A. erformance of the centr alized R-node deployment str ate gy first study the ef fects of the parameter 12 on the performance of the strate gy in terms of the probability that reliable path is found between an arbitrary source and destination that are separated by minimum hop-count (we shall refer to this probability as for con enience). It is desirable that should be small since otherwise, one ould ha to disseminate lar ge amount of control in- formation to enable node obtain this topology informa- tion. assume that the R-nodes are placed in accordance

to our centralized min-cut based strate gy Fig. sho ws that the strate gy is some what insensiti to the choice of (within reasonable set of alues that we can xpect to tak e). be more specific, the increase in when is in- creased from to is not significant (less than 0.1 in most cases). Since the comple xity of the min-cut algorithm in terms of running time is xoz| ?* we choose the lo wer alue i.e., ˆ‡ in all further studies. also point out that this means that only small amount of topological information is actually necessary for achie ving consid- erable impro ement

in performance (as to be seen later). Ne xt, we compare the performance results of the min- cut based centralized R-node deplo yment strate gy with those of se eral other R-node deplo yment strate gies. ‰ The longer the path, the higher the probability of its ailure. B Each node is assumed to kno the topology up to within hops of itself. 10 0.4 0.5 0.6 0.7 0.8 0.9 Number of hops of the shortest path between two nodes Probability 10% of all nodes are R−nodes =3, k=2 =3, k=3 =3, k=4 =4, k=2 =4, k=3 =4, k=4 Fig. 9. Comparison of the performance of the centralized R-node deplo yment

strate gy for arious alues of and Random str ate gy: strate gy in which the R-nodes are randomly deplo yed. De gr ee based str ate gy I: strate gy in which the R- nodes are placed in the proximity of nodes with the minimum de grees. ards this, we first sort the nodes in accordance with an ascending order in terms of their de gree. If there are R-nodes, the are placed in the vicinity of the first nodes in the or dered list. This strate gy appeared to be good choice initially since we ould xpect that the minimum de- gree nodes are the bottlenecks when attempting to find

multiple paths. De gr ee based str ate gy II: strate gy in which the nodes with the minimum de grees are identified first as in the pre vious strate gy; the R-nodes are placed in the proximity of the highest-de gree neighbors of these nodes (one neighbor for each node). do this since we recognize that the minimum de gree nodes may in act, be at the edges of the area that we con- sider and the bottlenecks may be due to the act that these nodes ha single link to the rest of the net- ork. Through this strate gy ,we attempt to mak such links reliable. From Fig. 10 and Fig. 11 we see

that the random R-node placement strate gy does not help much in find- ing reliable path between tw arbitrary chosen nodes when 10% of the nodes are R-nodes. It results in almost the same performance as that achie ed in case wherein there were no R-nodes. De gree based strate gy and de- gree based strate gy II can help in increasing ut the achie ed performance is still inferior as compared with the performance of the min-cut based strate gy by about 18% when ˆ‹ and by 25% when ¨Œ These com- parisons pro that the min-cut based R-node deplo yment
Page 11
10 0.1

0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Number of hops of the shortest path between two nodes Probability 10% of all nodes are R−nodes Min−cut based strategy Random strategy Degree based strategy I Degree based strategy II No R−nodes are deployed Fig. 10. Comparison of the performance of the arious R-node de- plo yment strate gies with O%P8R 10 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Number of hops of the shortest path between two nodes Probability 10% of all nodes are R−nodes Min−cut based strategy Random strategy Degree based strategy I Degree based strategy II No

R−nodes are deployed Fig. 11. Comparison of the performance of the arious R-node de- plo yment strate gies with O%PŽU strate gy is ery ef fecti and it of fers the highest alue of among all the schemes considered, especially when the number of deplo yed R-nodes is small. B. erformance of the distrib uted R-node deployment str ate gy The performance of the distrib uted R-node deplo yment strate gy without and with node mobility are studied ne xt. consider tw cases. In the first case, all the normal nodes are static, and only the R-nodes mo around and find their optimal

positions. Initially the R-nodes are scat- tered uniformly as well. In the later case, both the R-nodes and the normal nodes are allo wed to mo e. The random aypoint model is used to model normal node mobil- ity pattern. The speed of the normal nodes is uniformly distrib uted er JLKM The mo ving speed of the R- nodes is 10m/s, and the trajectories of these nodes are defined by the deplo yment strate gy This is in line with the requirement specified of R-nodes in sub-section -E. 10 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Number of hops of the shortest path between two nodes

Probability 10% of all nodes are R−nodes =3, distributed strategy without mobility =3, distributed strategy with mobility =3, centralized strategy =4, distributed strategy without mobility =4, distributed strategy with mobility =4, centralized strategy Fig. 12. Ef fects of mobility on the distrib uted R-node deplo yment strate gy Fig. 12 sho ws that the distrib uted deplo yment strate gy in the first case (without mobility), performs orse than the centralized deplo yment strate gy This is because, in the distrib uted strate gy the R-nodes do not ha central controller which can pro

vide global topology information. Based on the ailable local information that is dissemi- nated, the will ha to mo around and find their posi- tions. Some of the positions that the R-nodes choose may not be the optimal ones from the global point of vie Fur thermore, the netw ork topology changes with the mo e- ment of R-nodes, such changes mak it more dif ficult for the R-nodes to find the best positions. From Fig. 12 we see that the distrib uted strate gy per forms only little orse when there is mobility as com- pared with the case wherein there is no mobility (by about 5%

at most when ¸‹ ). The R-nodes can trace the topol- ogy changes in timely manner in spite of mobility and adapti ely modify their trajectories to find the best possi- ble positions. Thus, our distrib uted R-node deplo yment strate gy can be applied in practical mobile ad hoc net- orks, in which the normal ad hoc nodes are either static or ha pedestrian type motion. In this paper our objecti as to pro vide rob ustness to both intermittent (or short term) and long term node ail- ures in ad hoc netw orks. These ailures could be result of either ading, battery ailure or compromises.

The com- putation and use of multiple node-disjoint routes could potentially pro vide some tolerance to node ailures. proposed modifications to popular ad hoc routing pro- tocol OD to enable the computation of multiple node- disjoint paths without incurring the erhead generated by link-state routing methods. Our simulation results sho
Page 12
that the number of node-disjoint paths that can be found between source and destination depends on the density of nodes in the netw ork. Furthermore, we find that en at moderate node densities (a erage node de gree is 6.7), the

number of node-disjoint paths that may be found are ery limited (around if the distance on the shortest path be- tween the source-destinatio pair is 7). Thus, we infer that it is necessary to populate the netw ork with fe eliable nodes that are physically more sophisticated in terms of being capable of combating ading, possessing better bat- teries and physically more secure. These nodes which we call R-nodes are mainly used for creating reliable routing frame ork within the ad hoc netw ork. then attempt to address the question of where the R-nodes are to be posi- tioned within the ad hoc netw

ork and ho their trajecto- ries are to be controlled if notion of routing reliability is to be pro vided. define eliable path to capture the notion of routing reliability and aluate the performance of R-node deplo yment strate gies in terms of the probabil- ity that reliable path is found between source and destination. propose strate gy based on the random- ized min-cut algorithm. sho that our strate gy has the best performance in terms of the abo defined met- ric as compared with the other possible strate gies that we considered, and that it can cope with dynamic topology changes

due to lo mobility patterns. belie that the architecture proposed and de eloped, is necessary and is viable option for pro viding reliable routing frame ork in ad hoc netw orks. The authors ould lik to thank Chin ya Ra vishankar and Michalis aloutsos for their aluable suggestions and comments. [1] N.F Max emchuk, “Dispersity routing in store and forw ard net- orks, Ph.D. thesis, Univer sity of ennsylvania May 1975. [2] M.R. Pearlman, Z.J. Haas, .Sholander and S.S. abrizi, “On the impact of alternate path routing for load balancing in mobile ad hoc netw orks, Pr oceedings of the CM MobiHoc pp.

3–10, 2000. [3] A. Nasipuri, R.Castaneda, and S.R. Das, “Performance of multipath routing for on-demand protocols in mobile ad hoc netw orks, CM/Kluwer Mobile Networks and Applications (MONET) ol. 6, no. 4, pp. 339–349, 2001. [4] M.K. Marina and S.R. Das, “On-demand multipath distance ector routing in ad hoc netw orks, Pr oceedings of the Inter national Confer ence for Network Pr ocotols (ICNP) pp. 14–23, No 2001. [5] K. and J. Harms, “Performance study of multipath routing method for wireless mobile ad hoc netw orks, Pr oceedings of the IEEE Int’l Symposium on Modeling Analysis and Simulation

of Compute and elecommunication Systems (MASCO TS) pp. 99 107, 2001. [6] E. yanoglu, I. Chih-Lin, R.D. Gitlin, and J.E. Mazo, “Di ersity coding for transparent self-healing and ault-tolerant communi- cation netw orks, IEEE ansactions on Communications ol. 41, no. 11, pp. 1677–1686, 1993. [7] C.E. Perkins and E.M. Ro yer Ad hoc on-demand distance ec- tor routing, Pr oceedings of the IEEE orkshop on Mobile Com- puting Systems and Applications (WMCSA) pp. 90–100, 1999. [8] R. Motw ani and Ragha an, Randomized Algorithms Cam- bridge Uni ersity Press, 1995. [9] S. utukury and J.J. Garcia-Luna-Ace

es, “Mdv a: distance- ector multipath routing protocol, Pr oceedings of the IEEE IN- FOCOM pp. 557–564, 2001. [10] Zaumen and J.J. Garcia-Luna-Ace es, “Shortest multipath routing using generalized dif fusing computations, Pr oceedings of the IEEE INFOCOM pp. 1408–1417, 1998. [11] S.J. Lee and M. Gerla, “Split multipath routing with maximally disjoint paths in ad hoc netw orks, Pr oceedings of the IEEE ICC pp. 3201–3205, 2001. [12] D.B. Johnson, D.A. Maltz, and J. Broch, “Dsr: The dynamic source routing protocol for multihop wireless ad hoc netw orks, Ad Hoc Networking pp. 139–172, 2001. [13]

.D. ark and M.S. Corson, highly adapti distrib uted rout- ing algorithm for mobile wireless netw orks, Pr oceedings of the IEEE INFOCOM ol. obe, Japan, pp. 1405–1413, April 1997. [14] J. Raju and J.J. Garcia-Luna-Ace es, ne approach to on- demand loop-free multipath routing, Pr oceedings of the Interna- tional Confer ence on Computer Communications and Networks pp. 522–527, 1999. [15] S. R. Das, R. Castaneda, and J. an, “Simulation-based per formance aluation of routing protocols for mobile ad hoc net- orks, CM/Baltzer Mobile Networks and Applications pp. 179–189, 2000. [16] S.R. Das, C.E.

Perkins, and E.M. Ro yer “Performance compar ison of tw on-demand routing protocols for ad hoc netw orks, Pr oceedings of the IEEE INFOCOM ol. el vi Israel, pp. 3–12, 2000. [17] M. Ahmed, S.V Krishnamurthy R. Katz, and S. Dao, “T rajec- tory control of mobile gate ays that acilitate range xtension in ad hoc netw orks, Computer Networks ournal (COMNET) to appear [18] K. all and K. aradham, The ns Manual http://www .isi.edu/nsnam/ns/ns-documentation.html/. [19] S. i, Naldur g, and R. Kra ets, “Security-a are ad hoc routing for wireless netw orks, Pr oceedings of the 2001 CM Interna- tional

Symposium on Mobile Ad Hoc Networking Computing ol. Long Beach, California, pp. 299–302, 2001. [20] E.M. Ro yer and C.E. Perkins, re vie of current routing proto- col for ad-hoc mobile wireless netw orks, IEEE er sonal Com- munication Ma gazine pp. 46–55, April 1999.