A Distributed Routing Algorithm for Supporting Connection-Oriented Service in Wireless Networks with Time-Varying Connectivity Anastassios Michail Department of Electrical Engineering and Institute for Systems Research University of Maryland at College Park College Park, MD 07, USA tassos@eng.umd.edu Anthony Ephremides Department of Electrical Engineering and Institute for Systems Research University of Maryland at College Park College Park, MD 07, USA tony@eng.umd.edu Abstract We develop and simulate a distributed dynamic routing algorithm, capable of identifying paths for establishing and maintaining connection-oriented sessions in wireless communication networks which are characterized by frequent and unpredictable changes in connectivity. Our approach is a new protocol which runs atop a protocol for connectionless datagram service and establishes circuit routes for initial connection based on a mechanism of short packets exchange and on distributed information about availability of network resources. We explore the idea of predictive rerouting in that the algorithm takes advantage of the possibility to convert a connectivity change into a soft failure to maintain and re-route on-going sessions. The algorithm is simulated in Opnet and results show that the softening of link failures can improve performance as captured in terms of new call blocking probability and probability of forced termination of on-going sessions.. Introduction We consider the problem of dynamic routing in mobile wireless networks which are characterized by frequent and unpredictable changes in connectivity. We examine ad-hoc networks consisting only of mobile wireless communication nodes which move arbitrarily and do not pertain to any form of fixed network architecture, such as that of cellular networks. The existence of communication links between nodes depends on nodes location and distance, and on various physical layer factors such as transmission power levels, Prepared through collaborative participation in the Advanced Telecommunications/Information Distribution Research Program (ATIRP) Consortium sponsored by the U.S. Army Research Laboratory under Cooperative Agreement DAAL0-96--000. channel interference, antenna patterns and multipath propagation effects. Prior work [,,, 5] has focused on connectionless datagram service networks. The shift to connectionoriented type of service dictates the use of routing algorithms that can satisfy three main objectives: () discover routes as fast as possible, () establish connections between pairs of nodes through the discovered paths by reserving network resources and () react to connectivity fluctuations and maintain ongoing sessions, to the extent this can be feasible. Under these circumstances solutions to the routing problem should be provided by algorithms that execute in a distributed fashion without any need for global connectivity information. To these ends we have developed a fully distributed algorithm for identifying and maintaining paths between communicating pairs of nodes, based on a mechanism of short packets exchange. The starting point of our approach is a distributed routing algorithm for connection-less datagram service that was presented in [] and to which we will refer further on as CE algorithm. This algorithm assumes adequate bandwidth and an underlying access protocol, which is interference free. It relies on the execution by all nodes and separately for each potential destination of a two-phase procedure. The first phase ( query-reply message exchange) results in the establishment of a directed graph on a subset of the network that is rooted at the destination node and thus provides an initial set of routes. The second phase involves a similar structured exchange of control messages to react to the failure of an element of the previously established route and is intended to discover a new by-pass route. The algorithm adapts to large amounts of topological changes with no need for global topological knowledge, by building routes only as needed, instead of maintaining routes from all nodes to the potential destination nodes. Additionally, the extra routes built during
the query-reply phase increase protocol reliability. The most important properties of the CE algorithm encompass discovery of loop-free routes, deadlock-free operation and capability of detecting catastrophic network partitions. Shifting our focus to session-oriented service and, as is natural in this case, to the core of limited available bandwidth, we propose a new algorithm that establishes circuit routes for initial connection in the same manner of the CE data algorithm but that reacts differently, as it should, in the case of a link failure or connectivity change. The algorithm maintains the main properties of the CE routing algorithm but is modified to include in the message exchange additional information on the availability of network resources. In addition, in order to maintain on-going sessions in the presence of topological changes the algorithm takes advantage of the possibility to convert a wireless connectivity change to a soft failure. In the next section we present algorithm operation, starting with an overview and continueing with a definition of the network model, the router structure and a detailed description of the algorithm execution rules. In section we discuss performance evaluation and simulation results of the algorithm and we conclude in section with a short discussion of the main results and plans for future work.. Algorithm description.. Overview We consider multiple independently executing versions of the distributed routing CE algorithm, each one running for a specific destination node. The new algorithm for connection-oriented service runs atop the CE protocol and utilizes route information to establish new connections. The existing or new routes are explored in a hop-by-hop packet transmission mechanism in search of those paths which will guarantee admission of the request. As the CE protocol reacts to connectivity changes by reorganizing its routes, the overlaid protocol updates its information to be used either for attempts to accommodate new connection requests or to dynamically re-route on going sessions which experience quality degradation as a result of the changes in connectivity... Network model We model the network with a graph G =(N; L), where N represents the finite set of nodes and L the set of communication links. Each node i N has a unique node identifier (ID) and each link (i; j) L can be used either for one-way or two-way communication between nodes i and j, depending on the circumstances. All nodes are mobile and connectivity changes continuously, resulting in a time-varying set of communication links L. Each active link (i; j) L can either be undirected or directed. If the latter is true, then if the link is directed from i to j, node j is characterized as downstream (DN) neighbor of i. Similarly, if the link is directed from j to i node j is an upstream (UP) neighbor of i. An underlying link-level protocol is assumed which assures distributed knowledge of the changes in connectivity, in the sense that each node i is aware of all its adjacent nodes at all times, which are referred to as its neighbors. The set of the neighbors of node i varies also with time. We assume for simplicity that transmitted packets are received correctly and simultaneous two-way transmission over a link that would cause interference does not occur. The detailed mechanism of this link-level protocol operation is beyond the scope of this paper and will not be addressed at the present time... Router structure Every node has a fixed number of transmitter-receiver pairs (transceivers) used to set up communication links with other nodes. The number of transceivers that are in an IDLE state, varies dynamically with time depending on the traffic load and the average session duration. A necessary condition for a new connection to be established is that at least one transceiver is available at every node in the path to the destination. New connection requests are blocked when one or more nodes along the path do not have any of their transceivers idle. A pure first-come first-serve policy is assumed without considering any preemptive policies in which a high priority session can preempt an ongoing session of lower priority. For simplicity no priority is given to hand-off requests which have to compete against new connection requests in search of communication paths. A modified version of our algorithm in which hand-off requests are given priority over new call arrivals was presented in []. Sessions are distinguished by a unique ID, a numeric triple consisting of the source and destination IDs and a counter which is incremented by one for every new connection between the same source destination pair. Each node maintains a Connectivity Table with information on all sessions for which it serves as a source, destination or intermediate relay node. The connectivity table keeps a separate entry for each transceiver with the transceiver ID, the session ID, the incoming and outgoing link indices and the status of the communication transceiver... Algorithm execution Algorithm execution can be viewed as occurring in three logical phases, the Construction phase, the Maintenance phase and the Termination Phase, which execute simultaneously in a dynamic topology.
A. Construction phase During the construction phase mobile nodes desiring communication with other mobile users in the network, place connection requests which travel along paths provided by the underlying CE algorithm that terminate at the destination node. A request may be admitted and the session will be established if the chosen path can provide the sufficient network resources required for end-to-end flow of information. Without loss of generality we assume that at some point in time a node not adjacent to the destination node,, desires a connection. We also assume for simplicity that the QRY-RPY process (construction phase of CE) has already occurred and the part of the network under discussion can be represented by a directed acyclic graph V G rooted at the (figure (a)). Hence any node in this graph will always have at least one DN neighbor and by properties of the CE protocol there definitely exists at least one route initiating at any source node and terminating at the. A source node which desires a connection to the transmits a Connection-Request () packet along one of the existing DN links. If multiple DN links exist a decision over which link to transmit is made either upon information on the resources available along the existing paths or randomly, if no such information has been obtained. In particular the parameter for the selection of the DN link is the available number of transceivers along the outgoing paths and such information is collected during the algorithm construction phase by messages piggybacked in the transmitted acknowledgments. Unless it is the node, any other node receiving a temporarily reserves a transceiver (if at least one idle transceiver is available) and retransmits the to a DN node (see example in figure (b)(c)). If no idle transceiver is available at the time the is received, a Negative Acknowledgment control packet (NAK) is generated and sent back to the UP neighbor to indicate temporary lack of resources, and the particular link is blocked for future transmissions of the same. The NAK is a control packet generated by a node to indicate lack of resources to accommodate the request for a connection. Any node receiving a NAK attempts to retransmit the rejected to a different DN neighbor (if such a neighbor exists). To avoid multiple unnecessary attempts of transmitting over the same link, a link blocking rule is considered according to which reception of a NAK over a DN link automatically marks this link as Blocked for the specific. If a reaches the and the request is admitted, the destination node updates the corresponding entry in its connectivity table and transmits backwards to the source an Acknowledgment () control message (node in the example of figure (d)). Otherwise if the request cannot be serviced by the, a NAK is transmitted back to the (a) Initiate transmission (d) propagation CH (g) Link - fails Connection Handoff Request (j) Session handed-off to new path (b) propagation (e) propagation in reverse direction (cont) CH (h) CH reaches node (already in the path) (k) tears down session (c) propagation ) reached (f) propagation (session established) (i) propagation in reverse direction (cont) (l) tears down session Figure. Example of algorithm execution link over which the was received. messages are generated and transmitted by destination nodes, are destined to the source node of the and must follow the same path of the in the reverse direction (see example in Figure (d-f)), updating the connectivity tables of each node in the path. Link failures may destroy a path before the - phase is completed and this may result in reception of an over a DN link but for a session that has already been interrupted (before even acknowledged). A node receiving an updates the connectivity table by confirming the reservation and forwards the to the upstream node. A more enhanced version of the algorithm is when every is broadcast by a node to all its upstream neighbors. The advantage of this mechanism is based on a slight modification of the packet format to also carry the maximum number of available transceivers among all possible paths to the destination. This value is compared to the receiving node s idle transceivers and the minimum value replaces the entry in the packet field. Nodes do not anymore select randomly over which DN link to transmit a, but make their choice based on collected side information. This mechanism provides all nodes with the maximum amount of information but at the cost of outdated information in cases of large networks with high rate of topological changes, since by the moment the is received the information may already be inaccurate. It also results in a high number of control packets transmissions which could
slow down execution of the algorithm. B. Maintenance phase During the maintenance phase the algorithm reacts to connectivity changes that affect on-going sessions. In order to maintain connectivity for an additional amount of time, the algorithm takes advantage of the possibility to convert a wireless connectivity change to a soft failure, and search for a by pass route to handoff the on going session. We employ the notion of a soft failure (opposed to a hard failure which is a complete loss of connectivity) to characterize degradation in the link quality. A link which experiences a soft failure may still be used for transmission at some cost, for instance at a lower information rate, but the algorithm has the chance to locate and reserve a by pass route to be used by the failing session. The physical layer mechanisms during a soft failure could involve adaptive demodulation and efficient bit allocation algorithms which are beyond the scope of this paper. A node reacts to a soft failure by generating and transmitting a Connection-Handoff (CH) control message in search of by-pass routes. In general CH propagation follows the same rules as the with main difference that an may be generated either by the or also by any other node that is already in the path but is not affected by the link failure (see example in figure (g-j)). Accordingly or NAK propagation differs from the construction phase in that it ceases when the node which requested the hand off has been reached. Note that a node already in the path knows that a received CH is being served by comparing the session ID to in its connectivity table entries. Of course, when a hard link failure takes place connectivity is completely lost and the source node has to re establish the connection to the destination through the construction phase procedure. C. Termination phase During the termination phase, nodes clear entries in their connectivity tables that are no longer needed either because a connection was terminated or because it was interrupted due to a link failure or was handed off to a new path. When a session is completed the source node which initiated the call tears it down by generating a control message and broadcasting it along the session path to the (see example in figure (k-l)). Another situation where a packet is needed is when a link experiences a hard failure. In that case the two edges of the link lose communication with each other and therefore must notify the rest of the nodes located in the two resulting segments of the path to tear down the connection and clear the relevant entries in their tables.. Performance analysis We have simulated the system in Opnet Modeler, a discrete event simulator with the required features in modeling a distributed algorithm. In this section we highlight the main properties of the simulation model and present some initial simulation results... Mobile call model New call arrivals to the network are assumed to be Poisson with rate = request/min. The call holding time is exponentially distributed with mean = 0 sec. Each new call arrival is equally likely to arrive at any node as its source, and any of the remaining nodes is equally likely to be its destination... Connectivity model To capture node mobility and physical layer impairment, we employ a model of fixed-node topology but with dynamic link status. In particular, the source and destination of each call remain fixed as long as the call is in progress. Instead, the status of each link is assumed to be subject to dynamic (perhaps random) changes throughout the call duration. A simple three state probabilistic link status model is assumed. In particular, the possible states of the link status are FULL, HALF, or ZERO, corresponding to the status of the link being able to support transmissions at the full rate, half rate, or link out-of-service, respectively. Specifically, the transition from FULL to HALF state models a soft link failure, which parallels a link quality degradation situation, whereas the FULL to ZERO state transition corresponds to a hard link failure implying complete loss of connectivity. This model allows a unified treatment of node mobility and physical impairment in a flexible fashion. The difference from other commonly used two-state models [] is that we can model both a soft and a hard failure. We run simulations for different values of average holding times for each state and steady state distribution. The steady state distribution gives an indication of how much time on average each link spends at each case, while the holding times determine the average rate of topological changes, which gets higher when the holding times are decreasing... Simulation results Performance is measured end to end in terms of new call blocking probability and forced termination probability and compared under the assumptions of hard link failures only, against soft link failures. We consider a baseline topology of 0 nodes but examine two different connectivity senarios, topology (with a maximum of 7 links) and
topology (with a maximum of links). In all cases links are on average 95% of the time FULL and in the case of the state model % of the time under a soft failure situation whereas in the case of the state model 5% of the time under a hard failure. 0.8 0.7 0.6 Topology 0.8 0.7 Topology Probability 0.5 0. 0. Probability 0.6 0.5 0. 0. 0. Pf state model Pf state model Pb state model Pb state model 0. 0. Pf state model Pf state model Pb state model Pb state model 0 0 5 6 7 8 9 Average rate of link status change (changes/min) Figure. Results for Topology 0. 0 0 5 6 7 8 9 Average rate of link status change (changes/min) Figure. Results for Topology Figures and show the end to end probabilities of forced termination P f and of new call blocking P b as functions of the average rate of link status change l (in number of changes per link per minute) for the cases of topo and topo respectively. Note that P f increases as l increases since more sessions are interrupted. Clearly the use of the three state probabilistic model for the link status results in some improvement in P f. We also observe that P b decreases when l increases. In other words, as the link failures become more frequent, the new call requests have a better chance of being admitted end-to-end. The reason is that more capacity is made available at all nodes when more calls in progress get force terminated, so it is more likely for a new call to find capacity end to end. This indicates that a combination of blocking and forced-termination probabilities, rather than blocking alone, has to be considered in the end-to-end designs for such systems. Comparison of the results for the two topologies shows that the additional route diversity of topo results in a significant drop of P b whereas the improvement in P f is not that dramatic, mainly because of a limitation of the CE protocol which temporarily blocks some DN links in order to avoid formation of loops. Hence the temporary link blocking rule prevents our algorithm from taking full advantage of route redundancy.. Conclusion and future work We have presented an algorithm capable of supporting connection-oriented service in wireless all-mobile net- works. We have based our approach on a protocol for connectionless type of service to discover routes between source and destination pairs. The algorithm features a predictive re-routing scheme, where by the appropriate mechanisms wireless connectivity changes can be modeled as soft link failures and can maintain on-going sessions until bypass routes have been discovered. Simulation results show that performance improves when the three state probabilistic model is used. Our future work will focus on investigating the joint problem of routing and call admission control in order to achieve better improvement in performance as far as forced termination is concerned. We are also going to look into physical layer issues and power saving related procedures. References [] M. Corson and A. Ephremides. A distributed routing algorithm for mobile wireless networks. Wireless Networks, :6 8, 995. [] J. J. Garcia-Luna-Aceves and S. Murthy. A path-finding algorithm for loop-free routing. IEEE/ACM Transactions on Networking, 5:8 60, February 997. [] A. Michail, W. Chen, and A. Ephremides. Distributed routing and resource allocation for connection-oriented traffic in ad-hoc wireless networks. In Conference on Information Sciences and Systems, March 998. [] V. Park and M. S. Corson. A highly adaptive distributed routing algorithm for mobile wireless networks. In Proc. IEEE INFOCOM 97, Japan, April 997. [5] C. Perkins and P. Bhagwat. Highly dynamic destinationsequenced distance vector routing (dsdv) for mobile computers. In ACM SIGCOMM, October 99. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the Army Research Laboratory or the U.S. Government.