Control Message Reduction Techniques in Backward Learning Ad Hoc Routing Protocols

Transcription

1 Control Message Reduction Techniques in Backward Learning Ad Hoc Routing Protocols Navodaya Garepalli Kartik Gopalan Ping Yang Coputer Science, Binghaton University (State University of New York) Contact: Abstract Most existing wireless ad hoc routing protocols rely upon the use of backward learning technique with explicit control essages to route packets. In this paper we propose a set of techniques that can be applied in a backward learning routing algorith in order to iniize or even eliinate explicit control essages for route discovery, setup, and aintenance, while inially using iplicit data-like control essages that need no special processing. We also show that such an algorith does not need to prevent routing loops at all costs, such as by eans of extensive network-wide spanning trees in traditional LAN bridges, or destination sequence nubers in, or sourcerouting in. In fact, we prove that transient loops can be safely allowed to occur when a siple route refresh echanis is coupled with the use of packet identification field to effectively bound the lifetie of such transient loops without negatively ipacting the network perforance. Results deonstrate that even a routing algorith without explicit control essages can perfor copetitively in coparison to and protocols while significantly reducing the protocol coplexity. I. INTRODUCTION Existing ad hoc routing protocols [3], [5], [9], [7], [4], [6], [8], [2] set up and aintain loop-free routes by depending upon explicit control essage exchanges using an extensive control infrastructure. The use of explicit control essages enables these protocols to optiize the network perforance. For instance, the [9] protocol guarantees loop freedo through use of destination sequence nubers and relies upon explicit control essages such as RREQ, RREP, and RERR essages to construct and aintain routes. Siilarly, [7] uses explicit control essages, such as RREQ and RREP packets, to perfor source routing. While protocols-specific optiizations, such as piggybacking, can reduce the overhead of explicit control essages, significant protocol coplexity reains in place due to the need to aintain this control infrastructure and process the explicit control inforation. In this paper, we investigate whether one can design an ad hoc routing protocol that incurs very low control essage overhead and at the sae tie aintains copetitive network perforance. We propose a set of generic echaniss that can be applied to backward learning ad hoc routing algoriths in order to greatly reduce protocol coplexity. These techniques help to eliinate the use of explicit control essages as well as the need to aintain an overarching control infrastructure for route set up, aintenance, or loop prevention, and at the sae tie aintain a high network throughput. The key idea is to cobine the on-deand nature of ad hoc routing protocols with the siplicity of backward learning transparent bridges. While ost existing ad hoc protocols eploy the basic principles of backward learning, we exploit this principle without having to construct and aintain network-wide spanning trees as in wired LANs, or use destination sequence nubers as in [9], or resort to source routing as in [7]. Our results show that, even without explicit control essages, a protocol can deliver highly copetitive network perforance, if not better, in coparison to the popular and protocols. Using extensive siulations, we investigate the advantages and disadvantages of eliinating control essage overhead in order to achieve greater protocol siplicity. For clarity of exposition, we describe these techniques as ebodied in a proof-of-concept routing protocol called the Explicit Control Message Free () routing protocol. However, the proposed techniques can be readily adapted even to existing backward learning protocols, such as, to eliinate control essage processing. Just as in wired bridging and any wireless ad hoc protocols, nodes in protocol use backward learning to set up and adapt routes to different destinations on-the-fly. However, unlike conventional wireless ad hoc protocols, nodes iplicitly learn routes fro data packets theselves, rather than through explicit control essages. This works well for ost network counication sessions which involve bidirectional exchange of data, such as while using TCP. In the case of unidirectional traffic flow, selectively uses inial nuber of iplicit (or data-like) control essages that do not require any special processing in the network interior. At the sae tie, unlike wired bridges, also does not go out of its way to prevent routing loops at all cost but liits their perforance ipact by design, if and when such loops arise. Hence eliinates the overhead of constructing and aintaining an extensive infrastructure for loop prevention, such as network-wide spanning trees in wired-lan bridges, destination sequence nubers in, or source-routing in. We prove that any transient loops, which ay potentially arise, do not last beyond a bounded sall duration. Moreover, packets do not traverse around transient loops ore than once and thus do not lead to exponential packet stors. Additional features of include no need for proiscuous ode operation of wireless interfaces, no Hello essages between neighbors, and eliination of ARP (Address Resolution Protocol) request/reply essages. Thus, is copletely decentralized, self-starting, and autoatically adapts to the widely varying network conditions. II. PROTOCOL OVERVIEW nodes have two identities. First is at link level, such as layer-2 MAC address, which is visible only to iediate neighbors of the node. Second is at network level, such as

2 2 layer-3 IP address, which a node uses to identify and counicate with any other node (not necessarily a neighbor). A node does not need to have any prior knowledge of the identity, connectivity, or location, of other nodes in the network except, of course, the layer-3 IP address of the nodes with which it explicitly wishes to counicate. A node dynaically learns and unlearns the layer-2 MAC address identity of its neighbors only when it needs to counicate with/via those nodes. operates on top of link layer protocols, such as the 82. faily of protocols, which support link-level acknowledgent between neighbors to avoid unidirectional links. A network interface is not required to operate in proiscuous ode. Every node akes copletely local routing decisions that require no coordination with other nodes. is positioned as a thin protocol layer in the networking stack between the Layer-3 (IP layer) and the Layer-2 (MAC layer). Although forwarding is based on destination IP address, the routing is slightly different fro traditional IP routing. The next-hop address for a destination IP in the routing table is the layer-2 MAC address of the next-hop rather than its IP address. nodes do not care about the IP-to-MAC address appings of other nodes, except as needed to ake forwarding decisions. Routing Table and Data Packets: Each node aintains its forwarding knowledge in a soft-state routing table. Each routing table entry contains the following basic inforation: () Destination IP Address (D), (2) Next Hop MAC Address (), (3) A list of alternative next hop MAC addresses, (4) Valid flag (v), and (5) Expiration tie (e). Only routing entries that are valid can be used for packet forwarding. Routing entries can be invalid during the ties when a node is recovering after a failure or obility of a neighbor. A routing entry gets refreshed each tie a packet fro the corresponding destination D is received fro the next-hop MAC address. In other words, reverse traffic fro the destination D updates its forward routing entries at interediate nodes. This is unlike in the case of, where forward traffic towards D is used to refresh the forward routing entries. Expiration Tie field indicates the aount of tie left before an un-refreshed routing entry can be purged. We represent the relevant header fields of a packet by the quadruple [s, d, S IP, D IP ]. Here s and d represent the source and destination MAC addresses of iediate neighbors exchanging the packet and S IP and D IP represent the original source and final destination IP addresses. A represents a broadcast MAC or IP address. III. ROUTE DISCOVERY Figure illustrates the basic route setup echanis in. Source S initiates counication with destination D for the first tie by flooding its first data packet with header [S,, S IP, D IP ]. This data packet travels towards D, setting up routing table entries for S at interediate nodes. Any interediate node, that receives the packet fro S, processes the packet in two phases: the learning phase and the forwarding phase. Note that flooding a potentially large data packet could be ore expensive than flooding a saller (a) (b) S S S, *, S IP, D IP, *, S IP, D IP Fig.. W W, *, S, D IP IP Y,Y, S,D IP IP Illustration of route discovery echanis. control packet. In practice, ost counication sessions begin with an initial bidirectional exchange of sall packets (such as TCP SYN/ACK) which greatly reduces the flooding cost. A. Learning Phase In the learning phase, node learns about the direction in which node S is reachable. First consider the case in Figure (a) where is an iediate neighbor of S, but is not aware of this fact and does not have any prior routing inforation for S in its routing table. When receives the first broadcast packet with header [S,, S IP, D IP ], it iediately infers that a node with layer-3 IP address S IP can be reached in the direction of an iediate neighbor with layer-2 MAC address S. The node then enters this newly learned forwarding inforation in its routing table as a valid routing entry for S IP. Note that does not care that S IP and S are the sae neighbor s IP and MAC addresses respectively. All cares is that S IP is reachable in the direction of S. Consider the second case in Figure (b) when is not an iediate neighbor of S and does not have any prior routing entry for S IP. Rather receives a packet with header [W,, S IP, D IP ], forwarded fro another neighbor W, which does not know the next hop to D. Here will learn that S IP is reachable in the direction of W. Finally consider the third case, again in Figure (b), when node already has a valid routing table entry for S. Assue that a non-duplicate (i.e. first-tie) packet with header [W,, S IP, D IP ] is received fro a neighbor W, which already knows is a valid next hop for S IP. In this case the only new inforation that node can learn fro the packet is that S IP is still reachable via W and can refresh its routing table entry for S by resetting the expiration tie field. On the other hand, if the current next hop for S IP is not W then can infer that either the earlier known route to S IP has failed, or S has oved to a new location, or siply a better route to S ay have appeared depending on the specific link/path quality etric of interest. In this case, can choose to re-learn the new next-hop to S IP, as described later in Section IV. Another function of the learning phase is to learn alternative routes. Suppose an interediate node receives ultiple copies of the sae packet, which originated fro S, fro different neighbors. The first copy of the packet will be used by to learn the priary next hop towards S. Any subsequent copy of the packet will be recognized by as a duplicate (using echanis is described later in Section V). Before dropping the duplicates, will extract the source MAC address fro the duplicates as alternative next hop towards S. This alternative next hop can be used to replace the the priary Y D D

3 3 next hop, in case the priary link fails or the alternative has better link quality etric (such as bandwidth or noise). B. Forwarding Phase In the forwarding phase, node first checks if the packet is destined to itself, i.e. D IP = IP. If so, the packet is delivered to the IP layer (layer-3) within. If not, then deterines the next hop for destination D. If no valid routing entry exists for D IP then, as shown in Figure (a), rebroadcasts the data packet to its neighbors with a odified header [,, S IP, D IP ]. If already has a valid routing entry for D IP with the next hop neighbor Y then, as shown in Figure (b), forwards the data packet to Y with a odified header [, Y, S IP, D IP ]. Hence, if parts of the network already have valid routes to D then those routes are reused. If, for soe reason, the transission to next hop node Y fails, then invalidates the routing entry and attepts to send the data packet to one of its neighbors Z in the alternative next hop list of routing entry for D with a odified header [, Z, S IP, D IP ]. If transissions to none of the alternative next hops succeed, then siply re-broadcasts the packet with a odified header [,, S IP, D IP ]. To coplete the picture, by the tie the packet fro S reaches D, the interediate nodes in the network (including D) learn about the route to reach S. Siilarly, when the receiving user-level application at node D responds back with its own data to its peer application at node S, the interediate nodes (including S) learn about how to reach D as well. C. Handling Silent Endpoints Most real-world network sessions, such as TCP, engage in bidirectional exchange of data or control packets at application or transport level. In the unlikely event, where only one of the end-points actively sends traffic, the backward learning echanis would aintain routing inforation for only the active sender and not for the passive receiver. This creates an undesirable situation where all data packets to the silent receiver ight end up being flooded due to absence of routing inforation for the receiver. To rectify this situation, the layer at the receiver node D onitors whether packets fro S are being received for soe ACTIV ITY INTERV AL duration (say 8% of axiu route expiration tie) without the application at D sending any data back to S. If so, the layer at D transits a duy IP packet to S IP with zero byte data payload. As the duy IP packet fro D IP to S IP travels through the network, interediate nodes learn the routing inforation for D IP. Thus future data packets fro S to D are unicasted rather than flooded through the network. Since the duy IP packet is not eant for any specific application at S, it is silently discarded by the IP layer at S. layer at D stops sending duy IP packets once it observes that S is not sending any ore data packets. The duy IP packet is an iplicit control essage that is forwarded just like any other data packet without any special processing. Further, it is created only under the exceptional circustance when one of end-points is copletely silent for a long tie. layer at S also avoids initial broadcast of too any back-to-back packets fro S to D by buffering the packets following the first packet, until either the first data packet or duy IP packet is received fro D. IV. ROUTE MAINTENANCE A. Refreshing Soft Routing State Using Reverse Traffic The route refresh echanis eployed by is particularly different fro the refresh echaniss eployed by other protocols such as. Each routing entry in routing table has an associated lifetie, which indicates the tie duration after which the a non-refreshed routing entry will be deleted. Assue that a node s routing table contains a valid routing entry for a destination address D IP with next hop as node Y. The lifetie field of this routing entry is refreshed to a axiu lifetie value of MA ROUTE LIFETIME each tie a packet originating fro node D is forwarded by node Y to node. In other words, can reaffir the fact that D is still reachable via its neighbor Y when it receives a packet fro Y that has a source IP address of D IP. Note that reverse traffic originating fro D is used to refresh the forward routing inforation towards D. This in ore accurate than other routing protocols such as in which packets destined to D are used to refresh its routing inforation. B. Handling Mobility and Network Failures A failure is detected whenever a node doesn t receive a linklevel acknowledgent for its unicast transission. Consider Figure 2 in which there is a failure along a route fro S to D between neighbors and Y. When detects a failure during unicast transission of a packet to its neighbor Y, it first checks every routing entry that contains Y as the next hop address. If such a routing entry contains a list of alternative next hops, then the best alternative, as per the link etric used, is assigned as the priary next hop. Otherwise, the corresponding routing entry is arked as invalid. For exaple, in Figure 2, has four other neighbors L, M, N, and O, besides Y. Of these four, L happens to be the best alternative next hop towards destination D. Upon failure of counication with Y, first assigns L as the priary next hop and unicasts the packet for D to L. This is shown in case (a). If the counication with L fails as well and there are no ore alternative next hops, then floods the packet to all its neighbors hoping that at least one of the can reach D. This is shown in case (b). The scope of this flood is naturally liited by the packet s TTL value when it is received by. Further, nodes around the area that are affected by the failure can quickly re-learn about new routes to the destination D as soon as packets fro D flow through those nodes. V. DAMPING DUPLICATES WITHOUT A SPANNING TREE nodes do not construct or aintain any spanning trees. Instead they dap the propagation of duplicate packets by using the TTL and 6-bit Identification fields, which are carried inside the IP header of every packet. Every node reebers the Identification values of last k packets seen fro source node S. A packet is dropped only if its value is aong the last k reebered values or is saller than

4 4 L (a),l, S,D IP IP (b) Alternative next hop, *, S IP,D IP O M Failure N Next Hop Y D Fig. 2. Two cases when the counication link between and Y fails. Fig. 3. Coplete transient loops disappear in bounded tie. all of the. This technique detects duplicates and tolerates re-ordering of packets within a window of k packets. In evaluations, we observe that a window size as sall as k = 3 is ore than sufficient to detect and dap all duplicates even in the presence of packet reordering. The worst-case per-node state requireent in an N-node network would be O(kN). Also note that, use of the Identification field in is quite different fro the intricate echanis of destination sequence nubers in, which are used to altogether prevent routing loops as well as aintain route freshness. In contrast, the IP Identification field is used in only to detect and dap duplicate packets. VI. LIMITING THE IMPACT OF TRANSIENT LOOPS We define a coplete transient loop for a destination D as one in which each node in the loop has a valid routing entry towards the next hop node in the loop. A partial transient loop for a destination D is one in which at least one node in the loop does not have a valid routing entry for D. Consequently, such a node would broadcast the packets destined to D and the next node in the loop picks up the broadcast packet. No protocol can avoid partial transient loops whenever any interediate node needs to broadcast a packet to an unknown destination. does not attept to prevent transient loops at all costs. Rather, if and when transient loops are indeed fored, their ipact on network perforance is contained as follows. () guarantees that coplete transient loops last for less than a provably short bounded tie interval (We prove this clai below). (2) Since nodes perfor duplicate packet detection (Section V), no packet will traverse a loop ore than once. Rather a packet is discarded once it revisits any node. (3) Successive nodes decreent the TTL in each packet and discard the when TTL reaches zero. Theore : Lifetie of a coplete transient loop for any destination D is less than or equal to MA ROUTE LIFETIME. Proof: Consider the coplete loop in Figure 3 for destination D with k nodes N to N k, where the next hop for destination D at node N i is N i+ and at node N k is N. By definition, node D itself cannot be part of the coplete loop and hence is outside the loop. Furtherore, in order for node N i to aintain N i+ as its next hop entry for D, N i+ ust forward a packet with source IP address D IP to N i within MA ROUTE LIFETIME interval of its learning the next hop. Two cases arise. In the first case, node D does not send a packet through any node in the coplete loop for MA ROUTE LIFETIME duration. Thus N i+ cannot forward a packet with D IP as source IP address to N i. Consequently N i s routing entry for D will expire and becoe invalid, thus breaking the coplete loop within MA ROUTE LIFETIME duration. In the second case, node D does send a packet through node N i+ within MA ROUTE LIFETIME interval. Since node D is outside the loop, the packet fro D ust enter the loop fro soe external node N e. Without loss of generality, assue that N e is a neighbor of node N i+ that we are presently considering. Thus node D s packet arrives at N i+ fro N e. However, N e is different fro N i+2, which is the current next hop for D at node N i+ (since N e is outside the loop). Thus, as described in Section IV, N i+ will switch to aintenance ode to relearn its next hop for D by invalidating its routing entry for D via N i+2. This too would break the coplete loop within MA ROUTE LIFETIME interval. It is easy to see fro the above arguent that a coplete loop soon decays into a partial loop in which one or ore nodes do not have a valid routing entry towards D and broadcast the packets destined to D. However the partial loop also lasts only until the broadcasting node re-learns a new route to D, which would happen quickly if the broadcasting node lies along the path of data/duy-ip packets arriving fro D. Furtherore, D is guaranteed to send either a data or a duy IP packet within ACTIVITY INTERVAL. VII. PERFORMANCE EVALUATION In this section we show that, even without explicit control essages, the perforance of a backward learning algorith, such as, is highly copetitive, if not better, when copared to and protocols. Siulations were run using NS2 for different networks with 5 nodes. The nuber of connections were varied fro to between randoly selected node pairs. We used the obility scenarios fro []. Mobility speed of the nodes varies between to 2 eters/sec. Nodes ove within a rectangular field of 53 according to the Rando Waypoint Model. Each siulation is run for 9 siulated seconds. Pause ties were varied fro s to 9s. We used both TCP-based connections, (for bidirectional counication), and UDPbased Constant Bit Rate (CBR) connections (for unidirectional counication). The UDP CBR sources sent traffic at 4 packets/sec with 64 byte packets. Each data point is averaged over five siulation runs with a different rando seed to vary the connection establishent pattern. We count the duy IP packets as the control packets for. The MA ROUTE LIFETIME is set to 5 seconds and the duy IP packets are generated by the layer at a node after 4 seconds of silence on the part of the network application. For and, we count the Route Request (RREQ), Route Reply (RREP), Route Error (RERR), ARP requests/replies, and Hello essages as the control overhead. For, we do not count the source routing inforation carried in packet headers towards the control overhead. The control packets are counted on a hop-to-hop basis rather than on end-to-end basis. For fair coparison, in siulations we enabled the

5 5 Local Repair optiization (which attepts to repair a failed link locally before returning a RREQ essage to source) and the Link Layer Detection optiization (which suppresses the transission of Hello essages). Siilarly, the siulation of included the optiization of piggybacking RERR and RREP essages onto RREQ essages and the optiization of populating route cache by listening in proiscuous ode. We use three etrics : () is the ratio of the nuber of packets received by the destination to the nuber of packets sent by the source. (2) is the ratio of nuber of control packets to the nuber of data packets received by destination. (3) Nuber of control packets represents the total nuber of control essages transitted. A. Perforance with TCP Traffic : Figures 4, 5, and 6 show that perfors best in ters of values, followed by and. In Figure 4, as the pause tie increases, the value approaches (no packet loss). In Figure 5, as the nuber of connections increases, channel contention and packet losses increase resulting in a drop in. As in the previous case, perfors best, followed by and finally. Figure 6 shows a universal decrease in with increase in obility speed. The difference in between and is less than %. It can be seen fro these figures that reducing the protocol coplexity in does not autoatically translate into better delivery ratio () when copared against, because reduction in control overhead is offset to soewhat by increase in both data traffic. However, s delivery ratio reains copetitive with both and. : Figures 7, 8, and 9 show that aintains a low value of around.2 because it exploits the bidirectional nature of TCP traffic to aintain routing state without exchanging excessive control essages. and show higher value of for saller pause ties than for larger pause ties. experiences little variation in whereas variation is is large for both and. Figure 9 shows that increasing the obility speed of nodes results in greater increase in for than for and, because is unable to benefit fro the forward routing inforation carried in the reverse traffic. Nuber of Control Packets: Figure shows that nuber of control essages in is uch lower than and. For s pause tie, generates 4.5 ties fewer control packets than and 6 ties fewer control packets than. In Figure, we vary the nuber of TCP connections fro to 5, for a fixed 6s pause tie and fixed obility speed of 2/s. One can see that nuber of control packets generated by reains uch lower than and. As For 5 connections, generates around 22K control packets, generates around 25K control packets, while generates around 25K control packets. As with, Figure 2 shows that increasing the obility speed of nodes results in greater increase in nuber of control packets for than for and. B. Perforance with Constant Bit Rate (CBR) UDP Traffic In Figure 3, achieves an saller than both and for saller pause ties. In Figure 4, and achieve alost the sae which is saller than. When copared against results for TCP traffic, the above results show that, as expected, perfors better with bidirectional traffic, than with unidirectional traffic. Figure 5 shows that transits alost.5 control packets for each received data packet. Although, has approxiately.8 at s pause tie, it reduces as the pause tie increases. has a less than.2 at any given point of tie. Figure 6 shows that, as the nuber of connections increases, the value drops for. With ore nuber of connections, duy IP packets are suppressed ore often leading to lower value. On the other hand, value increases for both and. In Figure 7, generates around 4K control packets for s pause tie, around 75K, and around 2K. In Figure 8, uses far fewer control packets than others for large nuber of connections. Due to space constraints, we oit the obility speed variation results. VIII. CONCLUSION In this paper, we have proposed a set of techniques that can be applied to eliinate the use of explicit control essages in backward learning ad hoc routing protocols. The proposed techniques do not rely upon any extensive control infrastructure for loop prevention, such as spanning trees, destination sequence nubers, or source routing echaniss. They are particularly suited for routing bidirectional counication sessions with inial control overhead, and for unidirectional counication, use inial iplicit (data-like) control essages that require no special processing at interediate nodes. REFERENCES [] The onarch project. httpe// [2] I. Chakeres, E. Royer, and C. Perkins. Dynaic MANET on-deand routing protocol. IETF Internet Draft, October 25. [3] C.Perkins and P.Bhagwat. Highly dynaic destination-sequenced distance-vector routing (DSDV) for obile coputers. In ACM SIG- COMM 94, pages , 994. [4] R. Draves, J. Padhye, and B. Zill. Routing in ulti-radio, ulti-hop wireless esh networks. In Proc. of MobiCo, 24. [5] J. J. Garcia-Luna-Aceves and M. Spohn. Source-tree routing in wireless networks. In ICNP, 999. [6] Z. Haas, M. Pearlan, and P. Saar. The interzone routing protocol for ad hoc networks. IETF Internet draft, July 22. [7] D. B. Johnson and D. A. Maltz. Dynaic source routing in ad hoc wireless networks. In Mobile Coputing, volue 353. Kluwer, 996. [8] Y. Lee and G. Riley. Dynaic NIx-Vector routing for obile ad hoc networks. In Wireless Con. and Networking Conf., 25. [9] C. Perkins and E. M. Royer. Ad-hoc on-deand distance vector routing. IEEE Workshop on Mobile Cop Sys and Applications, 999.

6 vs 5 nodes - 3 TCP connections - 2/s vs # of Connections 5 nodes - 6-2/s vs Mobility Speed 5 nodes - 3 connections - 6 sec pause tie Fig. 4. vs for 5 nodes, 3 TCP connections, 2/s vs 5 nodes - 3 TCP connections - 2/s # of TCP Connections Fig. 5. vs TCP connections for 5 nodes, 6s pause tie, 2/s vs # of Connections 5 nodes - 6-2/s Mobility Speed Fig. 6. vs Mobility speed for 5 nodes, 3 TCP connections, 6s pause tie. vs Mobility Speed 5 Nodes - 3 Connections - 6 sec # of TCP Connections Mobility Speed Fig. 7. vs for 5 nodes, 3 TCP connections, 2/s..5e+5 e+5 5 vs 5 nodes - 3 TCP connections - 2/s Fig. 8. vs TCP Connections for 5 nodes, 6s pause tie, 2/s. 2.5e+5 2e+5.5e+5 e+5 5 vs # of Connections 5 nodes - 6-2/s Fig. 9. vs Mobility speed for 5 nodes, 3 TCP connections, 6s pause tie..5e+5 e+5 5 # of Control Pkts vs Mobility Speed 5 nodes - 3 connections - 6 sec pause tie Fig.. Control packets vs for 5 nodes, 3 TCP connections, 2/s vs 5 nodes - 3 UDP connections - 2/s Fig. 3. vs for 5 nodes, 3 UDP connections, 2/s vs # of UDP Connections 5 nodes - 6 Pause tie - 2 /s # of TCP Connections Fig.. Control packets vs TCP Connections for 5 nodes, 6s pause tie, 2/s vs # of UDP connections 5 nodes - 6 sec - 2/s # of UDP connections Fig. 4. vs Nuber of UDP connections for 5 nodes, 6s pause tie, 2/s. 2e+5.5e+5 e+5 5 vs 5 nodes - 3 UDP connections - 2/s Mobility Speed Fig. 2. Control pkts vs Mobility speed. 5 nodes, 3 TCP connections, 6s pause tie vs 5 nodes - 3 UDP connections - 2/s Fig. 5. vs for 5 nodes, 3 UDP connections, 2/s. 4e+5 3e+5 2e+5 e+5 vs # of UDP connections 5 nodes - 6-2/s # of UDP Connections Fig. 6. vs UDP Connections for 5 nodes, 6s pause tie, 2/s Fig. 7. Control pkts vs for 5 nodes, 3 UDP connections, 2/s # of UDP connections Fig. 8. Control pkts vs UDP Connections for 5 nodes, 6s pause tie, 2/s.