TRADING OFF CONSUMPTION OF ROUTING AND PRECISION OF MEMBERSHIP IN AD HOC NETWORKS

Transcription

1 TRADING OFF CONSUMPTION OF ROUTING AND PRECISION OF MEMBERSHIP IN AD HOC NETWORKS Juan Carlos García, Mari-Carmen Bañuls, Pablo Galdámez and Rubén Quintero Instituto Tecnológico de Informática, Universidad Politécnica de Valencia, Valencia (Spain). ABSTRACT The development of distributed applications and components for ad hoc networks rises specific questions. The need for energy efficient protocols is always one major concern in the design. Nevertheless, when several services are to be simultaneously provided, the choice of the most energy saving option for an individual component may be against the quality of other services. In this paper we focus on providing routing together with a global membership estimation. The former is a fundamental service in ad hoc networks. The latter will be required by any distributed application using consensus. Some proactive routing protocols, as OLSR, already maintain at each node information on the global composition of the system. With reactive routing algorithms, as DSR, a specific service must be used to monitor the global system composition. We perform a study of the joint performance of routing and failure discovery services in the case of OLSR, which provides both at once, and in the case of DSR routing complemented with gossip style failure detection. We compare the energy consumption of both approaches and show that the advantage of one routing service over the other may be compensated or inverted when bounds are imposed on the liveness of membership monitoring information. KEY WORDS ad hoc networks, distributed systems, wireless routing, energy consumption, consensus, group membership. 1 Introduction Mobile Ad hoc Networks (MANETs), or simply ad hoc networks, are wireless networks without a fixed infrastructure, that may be constituted by heterogeneous nodes with possibly more than one hop between them. Different network services are required to support the development of distributed applications over this type of systems. An outstanding case is that of routing protocols, that allow messages to be addressed to individual nodes, in spite of the specific problems due to mobility and limited node resources. Different specifications of routing protocols exist for ad hoc networks. Their energy consumption is one of the most decisive factors, but the choice among them for a given system will depend on the particular scenario. Nevertheless, a routing service may not fulfil all the requirements of certain distributed systems. In particular, an application could require information about which nodes are connected or disconnected in order to estimate the global composition of the system. An evident example is that of distributed consensus if some node may crash [1]. Since many routing protocols only maintain local information on neighbouring nodes, such estimation of global membership is in general not achievable by the routing service alone. There exist routing protocols that keep the global composition of the system at every node, e.g. OLSR (Optimised Link State Routing Protocol), and can provide membership information. As a drawback, OLSR is a proactive service, and thus it implies a certain sustained energy consumption. From the point of view of energy, a reactive routing algorithm optimised for lower consumption would be preferable. To our knowledge no reactive routing protocol provides every node with information about the global composition of the system. Thus an extra service should be implemented to perform the membership estimation. The best strategy to achieve this in a wireless ad hoc network is to build a gossip style protocol, that will indeed impose some extra energy cost, but which can be tuned independently of routing requirements. The goal of this paper is to analyse both approaches from the point of view of their performance regarding maintenance of global information and energy consumption. The rest of the paper is structured as follows. In section 2 we review some of the existing work related to routing services and gossip like protocols for ad hoc networks and their evaluation. Section 3 describes the protocols involved in both approaches under comparison, and section 4 collects the performed experiments and the results obtained. Finally, in section 5 we discuss the results and summarise our conclusions. 2 Related Work The multiple existing proposals for routing protocols in ad hoc networks can be classified in two groups, namely reactive (also called on-demand ) and proactive. Reactive algorithms, as DSR [2], AODV [3] and TORA [4], only operate when a new packet is to be sent and no calculated valid route exists from the sender to the receiver. The route search process introduces some delay in the delivery, but the overload due to network maintenance

2 messages is very low. On the contrary proactive algorithms, as OLSR [5] and DSDV [6], periodically exchange data about routes and system state in order to refresh the system topology. They imply a constant energy consumption by all nodes, and introduce a higher overload in the network. Different studies have been carried in order to characterise the performance of both approaches with respect to different parameters and in various working scenarios. In [7], the authors define the main characteristics and different parameters of evaluation of ad hoc networks. The energy consumption of different routing protocols is compared in [8] in terms of total, transmission, reception and packet type dependent energy consumptions, over 500x500 m 2 scenarios. [9] studies four protocols and compares their packet delivery ratio, routing overhead and path optimality using multi hop scenarios of 50 nodes with CBR traffic. OLSR is studied in [], which evaluates its control message overhead and loss of data packets over networks of more of 0 nodes. The second fundamental block of our work is the gossip style protocol. Epidemic or gossip protocols are widely used for spreading information in large distributed systems [11] without strict guarantees but high probability of success. Their use is typical for providing group communication services in wireless networks and has also been proposed for achieving computational tasks in large groups. As a membership estimation service, its use was first proposed in [12], whereas [13] defined a gossip like failure detection system in charge of monitoring changes to the network in a wired environment. The use of gossip protocols for routing in ad hoc networks was analysed in [14]. In that work several gossip variants are proposed and studied. We use one of them to adapt the gossip style failure detection in [13], so that instead of selecting a number of destinies for the gossipped message, each receiver probabilistically decides whether to forward data. This yields a gossip like protocol which provides each node with an estimation of the changes to the composition of the global network. 3 The Analysed Protocols OLSR routing and membership maintenance. OLSR [15, 5] is a proactive routing algorithm that maintains information on the global composition of the system at every node. Its proactive nature ensures that routes are available at any moment. OLSR is a optimisation of link state algorithms that reduces flooding overload only to a set of nodes called MPR (Multipoint Relays). OLSR handles two types of messages. Hello messages are used to discover links between neighbours and their unidirectional or bidirectional character, and TC messages to build the overall network topology. Hello messages are exchanged periodically between neighbours. Each node selects its set of MPR nodes from received Hello messages and periodically broadcasts a TC message to the entire network with part of the received information. This message is only forwarded by MPR nodes and it is used by every node to update the global composition of the system. The behaviour of OLSR is determined by a number of parameters, among which the most relevant for our comparison are H i, the interval between Hello messages; T C i, the interval between TC messages; H t, the time validity of the last Hello message (H t H i ); and T C t, the time validity of the last TC message (T C t T C i ). DSR with Gossip Style Failure Detection. DSR (Dynamic Source Routing) [2] is one of the most widely used reactive algorithms. When a packet is to be sent to an individual node, and if no pre-calculated route to the receiver exists, the sender produces and broadcasts a Route Request message, containing its identity and that of the destination, a unique message identifier and a path list initially holding the sender alone. The neighbouring nodes that receive the Request add their identifier to the route path list and broadcast the modified message. After receiving it, the target replies to the source with a Route Reply message that contains the path followed by the message from the sender. Since Route Requests are only issued when there are messages to be sent, no constant overhead is imposed by DSR, at the price of a certain delay when delivering a message to a node for which no route is yet known. On the other hand, as a node N i only holds information on nodes for which a route including N i exists, there is no guarantee that the global composition of the system is locally available. Liveness is ensured by confirming the reception of data packets at each hop. If a node misses such confirmation from the next node after a number of retransmissions, it generates a Route Error message to notify the original source about a broken link in the route. For the maintenance of global membership information we adopt a gossip like protocol from [13], slightly modified to adapt it to the wireless scenario. Each node maintains a table of connected nodes, with their state, a (heartbeat) integer value and the last time that value was updated. Every T g seconds each members increments its heartbeat and, with probability P hb, broadcasts its local table. The receivers merge the incoming information with their local table, updating entries for which newer heartbeat values have arrived, and setting the state of possibly failed nodes. Then they rebroadcast the updated table with probability P s. When the heartbeat of a certain node is older than a predefined time T f, the node is considered to have failed, its state is changed in the local table and after a longer period T cleanup, it is definitely removed from the table. When a node joins the network its table keeps only its identifier and the initial heartbeat. During a fixed time T s the node waits to receive an extended table from an existing member. If that does not happen, it broadcasts its own and enters normal operation. The main parameters controlling the gossip algorithm are T g, the periodicity of heartbeat increment; T f, the maximum allowed age of information on a particular node before setting its state to failed (generally T f 2T g ); T s, the length of the initial waiting period when

3 joining the network; P s, the probability of broadcasting updated information after reception of new data; and P hb, the probability of broadcasting the updated information at each Gossip period. 4 Measurements and Results Our comparison between both approaches is centred on two aspects. On one hand, we are interested in the quality of the obtained global information. On the other, we must evaluate the energy consumption in any case. In order to quantify the liveness of global membership information we measure the time it takes for all nodes in the system to reach a common knowledge about the connected members after a change has happened. This is 1 done in two different scenarios. In the first one, one joins, the N-th node joins the group when the remaining N 1 are in a stable state (all of them having the same correct knowledge about present members). We measure the time elapsed from this join until the last node knows the full group composition,. (j) In the second scenario, one leaves, one node is switched off when all N nodes were in the stable state, and we measure the time it takes for the remaining members to notice the failure,. (f) In each execution we also measure the total energy consumed by each node as a function of time. This is found to fit to a straight line, whose slope determines the consumed power, de/dt, approximately constant in every execution. We average this quantity over all the nodes in the network, to obtain a measure of the energetic performance of the protocols. Description of the Experiments. Our simulations were done using ns [17]. The Network Simulator, ns2, includes a number of protocols for wireless networks, among them, DSR. Although the employed release of the simulator does not include a version of the OLSR protocol, there exist implementations developed by third parties for their use with ns2. We used the freely downloadable one from NRL Protean Group [18]. The gossip protocol described in sect. 3 was explicitly implemented to be included in the simulator. This implementation is available at [19]. We performed a series of experiments to measure, (j) (f) and de/dt as a function of the relevant protocol parameters. Each experiment consisted of a set of 20 simulations with varying network topology, whose results were later averaged. Topologies were constructed by randomly locating the N nodes on a square grid of dimension m 2 in the case of single hop and m 2 for multihop scenarios. Since the range of radio signal assumed by ns2 is 250 m, in order to ensure connectivity the maximum allowed distance between a node and its nearest neighbour was 60 m for single hop, and 225 m for multihop settings. Each simulation was run for the one joins 1 This kind of measurements have been previously used to characterise the behaviour of gossip algorithms in wired networks [16]. scenario to determine (j) and de/dt and for the one failures scenario to determine. (f) The experiments were repeated for various values of the total number of nodes, from N = 4 to N = 40, for both proposed architectures, gossip with DSR (GDSR) and OLSR. The simulation time was 300 s for single-hop scenarios and 00 s for multi hop networks, after the state of the nodes had stabilised and they shared the correct knowledge about the initial composition of the network. The power consumption for reception and transmission were fixed to 0.2 mw. The basic factors that determine the power consumption and the liveness of global information in each case are the periodicity of information broadcasting and the lifetime of local information. The performance of both proposed approaches has to be compared when those factors are comparable. In the case of GDSR the relevant parameters are the gossip interval T g, the time to detect a failure, T f and the probabilities of forwarding local or received information, P hb and P s. In our simulations we considered P hb = P s, so that we had only three free parameters. In OLSR, the periodicity of information refreshing is determined by H i and T C i, whereas the time to detect a failure is governed by H t and T C t. We took H i = T C i and H t = T C t in our experiments, so that only two independent parameters remained. The role of H i and T C i in OLSR is equivalent to that of T g /P g in the gossip failure detection, namely the period between local broadcasts from each node. 2 On the other hand, H t and T C t determine the expiry of information, as T f does in the case of the gossip algorithm. This consideration allowed us to establish comparable sets of parameters in both protocols. We performed a group of experiments with fixed P s = 0.06 and T g between 1 and 5 s in GDSR, and with the corresponding values H i = T g /P g in OLSR. In the one joins scenario we fixed T f = 2T g and H t = 6H i. In one leaves a more realistic T f = 0.7T g /P g was chosen, as this upper bounds the probability of detecting a fake failure due to missing gossip messages, and H t = 1.5T g /P g (as the implementation of OLSR requires H t > H i ). Results. Figs. 1 and 2 show in a logarithmic scale the dependence of (j) with T g for both approaches in single and multi hop scenarios, respectively. Notice that the value of T g determines that used for H i, the relevant OLSR parameter, through H i = T g /P s. In the plots (j) shows a linear dependence with broadcast periodicity, less pronounced in the case of GDSR, and with little sensitivity to the number of nodes in the network, N, specially in the case of OLSR. As expected, the best time to reach agreement when a new member joins is found for GDSR, with (j) at least one order of magnitude lower than in the case of OLSR. GDSR shows better behaviour regarding the agreement time, since its operation is optimised to react when 2 Equivalently we may say that in OLSR each node broadcasts its information 1/H i times per unit time, whereas in the gossip algorithm it does so P g/t g times per unit time.

4 (s) (j) 3 GDSR,N=12 2 GDSR,N=12 (P=0.03) (P=0.03) (s) (f) 120 GDSR,N= Figure 1. T (j) agr as a function of T g in a single hop scenario. Figure 4. T (f) agr as a function of T g in a multi hop scenario. (s) (j) 3 2 GDSR,N=12 GDSR,N=12 (P=0.03) (P=0.03) de/dt(w) -2 GDSR OLSR GDSR (P=0.03) GDSR (P=0.01) N Figure 2. T (j) agr as a function of T g in a multi hop scenario. Figure 5. Power as a function of N in a single hop scenario. (s) (f) 60 GDSR,N=12 50 de/dt(w) -2 GDSR OLSR GDSR (P=0.03) GDSR (P=0.01) N Figure 3. T (f) agr as a function of T g in a single hop scenario. Figure 6. Power as a function of N in a multi hop scenario.

5 new nodes appear. It is highly probable that a new node receives the composition of the network during its initial waiting time T s and it takes only one flood to spread the complete information to the entire network. Receiving such initial message is more likely for larger N, specially in the case of a single hop. In multi hop scenarios, higher agreement times are obtained (although still far lower than those for OLSR) as the information needs to go through a number of hops, and for constant N the probability of having a gossip within one transmission range is lower. The time to reach agreement in the one leaves scenario, in figs. 3 and 4, shows in most cases a linear dependence with T g. For GDSR it is directly governed by the time of validity of local information, given by T f. It would thus be possible to tune the reaction time, at the expense of tolerating some spurious failure. In the single hop scenario OLSR manages to reduce (f) even under the value for GDSR. This is due to its capability of monitoring link quality, discarding links that do not show activity for long enough periods. The effect shows when the periodicity is low enough, but it is not relevant in the multi hop scenario. Different results are found for energy consumption. Figs. 5 and 6 show in a logarithmic scale the average consumed power with fixed T g = 1 s, as a function of the number of nodes in single and multi hop scenarios. Contrary to what happens with liveness, the energetic performance of OLSR exceeds that of GDSR when comparable parameters are used for both protocols (starred and circle data series). The average consumption rate is lower and scales better with the number of nodes for the OLSR approach both in the case of single and multi hop scenarios. Summarising the results above, GDSR yields much better liveness for changes detection, whereas OLSR energy consumption is lower. Yet, the magnitude of the difference in the values obtained for (j) is such that one may try to trade off some liveness for energy saving. Thus, we performed a second set of experiments with GDSR, in which we limited the liveness by taking P s = 0.03 while keeping the remaining parameters unchanged. The results of such tests appear also in the figures above. In figs. 1 and 2 the effect of lowering P s in the liveness of information about joined nodes is shown (series with cross-shaped marker). We see that the result is worse than the one obtained with P s = 0.06, as expected, but it is still much better than the one obtained with OLSR. In 5 and 6 we observe the effect of this change in energy consumption. In the single hop scenario, the OLSR approach is still more efficient, except for the lowest values of N. However, in the case of multi hop scenario, GDSR with P s = 0.03 yields a better result for all system sizes, between 4 and 40 nodes. A similar improvement is achieved in the single hop experiments by further reducing the liveness to P s = 0.01, whereas the time of agreement is still better than for OLSR. 5 Conclusions We have analysed the joint performance of routing and membership estimation, two services that are to be commonly used in MANETs. Two approaches have been proposed, namely a single protocol (OLSR), which provides both of them at once, and a combination of two specialised algorithms (GDSR). OLSR has the advantage of needing a single component, with the implicit optimisations of the OLSR protocol regarding flooding in a multi hop network. GDSR offers higher flexibility, as the parameters of the gossip like protocol can be independently tuned to obtain the desired balance between energy consumption and liveness of information, whereas the reactive behaviour of DSR does not impose a sustained flow of control messages. We have simulated both approaches with comparable configurations. The results show that depending on the characteristics of the system (size, single or multi hop character), and on the requirements on the composition information (degree of liveness), either alternative may be preferable. GDSR may be easily tuned to provide very fast detection of joined nodes, whereas OLSR is more efficient in the consumption of energy. In order to get a competitive energy result with GDSR, liveness must be relaxed. In multi hop scenarios the balance is more easily achievable, and GDSR may provide better results in both time of agreement and consumed power. The performance of this approach could be further optimised if an on demand approach is adopted to compose the global view of the network only when necessary [20]. Furthermore, we have also run tests in which a CBR application is run on top of either routing service, randomly sending messages to all known members of the network at a total rate of msg/s. The results indicate that the difference between OLSR and GDSR consumption reduces, but it is still possible to get better energetic results from GDSR with lower values of P s. Previous works studied the performance of routing protocols from the point of view of throughput or energy consumption, but to our knowledge no analysis was done of how other services could affect such results. We have explicitly shown that if DSR is chosen as a more energy efficient routing protocol, and some estimation of the network global composition is required by user applications, the component added to provide such information may neutralise the presumed saving, and even result in a more consuming system. Thus it is not enough to study the performance of some individual component to characterise the overall performance of the system or to optimise its energy consumption. In our opinion, it is more sound to carry out a study of the performance of the set of support services that are to be deployed, in order to understand their interaction and to reach a trade off between consumption and quality of service in a global manner. Our tests have been made using DSR and OLSR, but some of the arguments apply to reactive and proactive algorithms in general. Reactive algorithms offer the advantage of not forcing a constant traffic overload, whereas proactive

6 algorithms offer lower delays. However, when combining a routing protocol with other distributed service, we may find situations in which either approach exceeds the other s performance depending on the required QoS of the added component. If additional services are required by the applications, more components will affect the energetic performance and must be taken into account in this analysis. One major question concerning wireless networks is that of securing communications. As a future work we plan to integrate securing mechanisms into the protocols analysed so far to study their effect on the previous results. Acknowledgements. Partially supported by the Spanish Research Council (CICYT), under grant TIC C02-01, by Valencian Government under grant GV05/259 and by the Polytechnic University of Valencia, under grant The authors are grateful to Prof. Pietro Manzoni for fruitful discussions. References [1] Michael J. Fischer, Nancy A. Lynch, and Michael S. Paterson. Impossibility of distributed consensus with one faulty process. J. ACM, 32(2): , [2] D. Johnson, D. Maltz, and J. Broch. DSR The Dynamic Source Routing Protocol for Multihop Wireless Ad Hoc Networks, chapter 5, pages Addison-Wesley, [3] Charles E. Perkins and Elizabeth M. Royer. Ad-hoc on-demand distance vector routing. In WMCSA 99: Proceedings of the Second IEEE Workshop on Mobile Computer Systems and Applications, page 90, Washington, DC, USA, IEEE Computer Society. [4] Vincent D. Park and M. Scott Corson. A highly adaptive distributed routing algorithm for mobile wireless networks. In INFOCOM 97: Proceedings of the IN- FOCOM 97. Sixteenth Annual Joint Conference of the IEEE Computer and Communications Societies. Driving the Information Revolution, pages , Washington, DC, USA, IEEE Computer Society. [5] P. Jacquet, P. Mühlethaler, T. Clausen, A. Laouiti, A. Qayyum, and L. Viennot. Optimized link state routing protocol for ad hoc networks. In Proceedings of the 5th IEEE Multi Topic Conference (INMIC 2001), [6] Charles Perkins and Pravin Bhagwat. Highly dynamic destination-sequenced distance-vector routing (DSDV) for mobile computers. In ACM SIG- COMM 94 Conference on Communications Architectures, Protocols and Applications, pages , [7] S. Corson and J. Macker. Mobile Ad hoc Networking (MANET): Routing Protocol Performance Issues and Evaluation Considerations, [8] Juan-Carlos Cano and Pietro Manzoni. A performance comparison of energy consumption for mobile ad hoc network routing protocols. In MASCOTS 00: Proceedings of the 8th International Symposium on Modeling, Analysis and Simulation of Computer and Telecommunication Systems, pages 57 64, Washington, DC, USA, IEEE Computer Society. [9] Josh Broch, David A. Maltz, David B. Johnson, Yih- Chun Hu, and Jorjeta Jetcheva. A performance comparison of multi-hop wireless ad hoc network routing protocols. In Mobile Computing and Networking, pages 85 97, [] Anis Laouti, Paul Mühlethaler, Abdellah Najid, and Epiphane Plakoo. Simulation results of the olsr routing protocol for wireless network. Technical Report RR-4414, INRIA, March [11] Patrick T. Eugster, Rachid Guerraoui, Ane-Marie Kermarrec, and Laurent Massoulié. Epidemic information dissemination in distributed systems. IEEE Computer, pages 60 67, May [12] Richard A. Golding. A weak-consistency Architecture For Distributed Information Services. Computing Systems, Technical Report UCSC-CRL [13] R. Van Renesse, Y. Minsky, and M. Hayden. A gossip-style failure detection service. In Proceedings of Middleware 98, pages IFIP, The Lake District, UK, [14] Zygmunt J. Haas, Joseph Y. Halpern, and Erran L. Li. Gossip-based ad hoc routing. In INFOCOM, [15] P. Jacquet et al. Optimized link state routing protocol (olsr). Internet Draft (draft-ietfmanet -olsr-09.txt), INRIA Rocquencourt, April [16] Mark W. Burns, Alan D. George, and Bradley A. Wallace. Simulative performance analysis of gossip failure detection of scalable distributed systems. Cluster Computing, (2): , [17] The VINT project. The NS2 manual. Technical report, ISI, [18] NRL Protean Group. [19] [20] Mari-Carmen Bañuls and Pablo Galdámez. Ondemand membership for energy-aware networks. In First International Workshop High Availability of Distributed Systems HADIS 2005, 2005.