Adaptive Parameter Estimation Based Congestion Avoidance Strategy for DTN

Transcription

1 Proceedings of the nd International onference on oputer Science and Electronics Engineering (ISEE 3) Adaptive Paraeter Estiation Based ongestion Avoidance Strategy for DTN Qicai Yang, Futong Qin, Jianquan Bi, Qiubo Xu No Troops of PLA Luoyang, hina Abstract Delay Tolerant Networks are characterized by interittent connectivity, long delays and often constrained bandwidth. Delay Tolerant Networks often use ulti-copy routing schees for essage transission. Multi-copy routing schees, such as Epideic Routing, usually lead to congestion probles. In this paper, we propose an Adaptive Paraeter Estiation based congestion Avoidance strategy () for Delay Tolerant Networks. The strategy includes buffer anageent policy and packet receiving policy. The buffer anageent policy is designed for iproving the delivery ratio, and the packet receiving policy is designed for reducing the network overhead. The siulation results show that the strategy significantly iproves the delivery ratio and reduces the network overhead, copared to the default Drop-First- Received strategy. Keywords-DTN; congestion avoidance; buffer anageent; packet dropping policy I. INTRODUTION Delay Tolerant Networks (DTNs) [] are characterized by interittent connectivity, long delays and often constrained bandwidth. DTNs represent a broad class of networks such as interplanetary internets, under water sensor networks, and vehicular networks. The characteristics ake routing and congestion control becoe challenges. Then, the DTNRG [] was created by the IRTF to research DTN, and a new protocol naed Bundle Protocol (BP) [3] was introduced to face the challenges. The BP includes a store and forward transfer echanis to eet these challenges. Many new technologies are studied in DTNs, such as new routing protocols, new congestion control strategies. DTNs often use ulti-copy routing schees for essage transission. Multi-copy routing schees, such as Epideic Routing [4], usually lead to congestion probles. So congestion control strategies can significantly iprove the overall perforance of DTNs. In this paper, we propose the Adaptive Paraeter Estiation based congestion Avoidance strategy () for DTNs. The strategy includes both buffer anageent policy and packet receiving policy. We ipleent the strategy in the siulator ONE [5] for evaluating its effects. The siulation uses Rando Way Point odel (RWP). The strategy is copared with the ost used (Drop the First Received packet in the buffer) strategy. To analyze the general perforance of the strategy, we evaluate four perforance etrics: delivery ratio, average latency, average hop count and overhead ratio. The siulation results show that the strategy significantly iproves the delivery ratio and reduces the network overhead. The rest part of this paper is organized as follows: section II presents the related works. In section III, we discuss the strategy. Section IV displays siulation results and analyzes the perforance of the strategy. Finally, we ake a short conclusion in section V. II. RELATED WORK The DTN characteristics ake routing and congestion control becoe challenges, such as interittent connectivity, long delays and often constrained bandwidth. DTNs often use ulti-copy routing schees for essage transission. These ulti-copy routing schees often lead to congestion probles, because there are too any copies in the network. Therefore, we need to study congestion control strategies to iprove the overall network perforance. There are about three categories of congestion control strategies in DTNs. The first category is congestion avoidance strategies. These strategies generally restraint the essage transission, to decrease the nuber of essage copies in the network. Paper [6] discusses a Token Based ongestion ontrol echanis. Network nodes ust possess a token in order to inject essages into the network. Tokens are initially uniforly distributed but thereafter ove randoly throughout the network. This echanis results in a reduction in the aount of network transit ties. In paper [7], the authors propose a new congestion avoidance echanis based on path avoidance for DTN. The echanis optiizes the anageent of node storage and divides the node state into noral state, congestion adjacent state, congestion state. Based on different node states, separate strategy is eployed correspondingly to avoiding congestion. The second category is flow control strategies. These strategies generally restraint the essage flow between neighbors. In paper [8], a hop-by-hop Local Flow ontrol over InterPlaNetary networks (IPN) is discussed. Its ain contribution is the introduction of an asynchronous congestion control echanis on the network that not presented in the traditional DTN protocol. Through the hopby-hop local flow control, this echanis tries to resolve buffer overflow, link saturation and useless retransission probles in order to better utilize the overall syste. The third category is buffer anageent strategies. When the network becoes congested, different buffer scheduling strategies are deployed for iproving the overall perforance. In paper [9], authors propose an optial joint scheduling and drop policy. This policy ais to optiize

2 Proceedings of the nd International onference on oputer Science and Electronics Engineering (ISEE 3) different perforance etrics, such as the average delivery rate and the average delivery delay. Paper [] introduces a novel congestion control strategy, called average forwarding nuber based on epideic routing (AFNER). According to the forwarding nuber queue, the strategy deterines the packet forwarding sequence. In this paper, our strategy includes both buffer anageent policy and packet receiving policy. The buffer anageent policy is designed for iproving the delivery ratio, and the packet receiving policy for reducing the network overhead. III. THE DESIGN OF THE STRATEGY In Delay Tolerant Networks, the average latency of network and average hop count are both finite and each is around soe value. Thus, when the network becoes congested, we consider dropping packets with ore hops and larger latency first. This is the idea of the buffer anageent policy. But this policy usually cannot observably reduce the network overhead, so we develop the packet receiving policy for reducing ost useless transission in the network. Packets with ore hops and larger latency will be received in very low probability. The buffer anageent policy and the packet receiving policy for our strategy. A. The Buffer Manageent Policy As entioned above, when the network becoes congested, we consider dropping packets with ore hops and larger latency first. We use the inforation contained in the node itself to estiate the average latency and the average hop count of network. Then, we use the estiation values to calculate the priority weight of each packet. Packets with ore hops and larger latency correspond to saller priority weights. When the network becoes congested, we drop the packet with sallest priority weight first. For node i in the network, i is the nuber of packets in its buffer, T ij is the living tie of packet j in node i, and H ij is the hop count of packet j in node i. The estiation values and the priority weight are calculated as follows: ) The average latency estiation i T= i Tij () i j= is the latency estiation constant, usually ~4. ) The average hop count estiation i H= i Hij () i j= is the hop count estiation constant, usually ~4. 3) The priority weight function for node i Wi = α exp( Tplt /( Ti)) (3) + ( α) exp( H / H ) p ni i α is the latency ter coefficient, α (,). T plt is the current packet living tie. H p is the current packet hop count. ni is the estiation for the average nuber of neighbor nodes. According to [], for an Ad Hoc network with N nodes, the average hop count is usually around N, and the average nuber of neighbor nodes is around ln( N ). Thus, we get the estiation for ni : ni = ln( H i) (4) The coefficients, and α do affect the perforance of. These coefficients can be set to different values for different scenarios. Usually, we set =.5, =.5 and α =.5, if there is no special request. Though the forulas, we can find that the calculating process just use the inforation contained in the node itself, so this policy can be deployed in various scenarios. B. The Packet Receiving Policy The buffer anageent policy usually cannot observably reduce the network overhead, so we develop the packet receiving policy for reducing ost useless transission in the network. Packets with ore hops and larger latency, which correspond to saller priority weights, will be received in lower probability. Thus, the packet receiving policy directly reduces the network traffic. The designed packet receiving policy is described as follows: ) If the buffer using rate is lower than r t (< r t <), we consider that the network is not in the congestion state, and there is nothing ore to do. ) If the buffer using rate equals to r t or is larger than r t, we consider that the network is in the congestion state. If the destination of the coing packet is the current node, this packet ust be received. 3) If the buffer using rate equals to r t or is larger than r t, we consider that the network is in the congestion state. If the destination of the coing packet is not the current node, The following echanis will work: We calculate the average priority weight W AVG and the axial priority weight W in the current node. Then we MAX calculate the priority weight W of the coing packet (in the W calculation, the hop count H p plus ). We copare W with W and AVG W : MAX a) if W < W, the current node will refuse to AVG receive the coing packet; b) if W > W, the current node ust receive the MAX coing packet; c) if W AVG W W MAX, the current node will receive the coing packet in probability p:

3 Proceedings of the nd International onference on oputer Science and Electronics Engineering (ISEE 3) W WAVG p = (5) WMAX WAVG When receiving the new packet, if there is not enough free space for the new packet, the current node will free enough space according to the buffer anageent policy. Usually, r t has little effect on the network perforance, and it is suitable to set r t =.8. Notice that, in step 3), when W < W AVG, the current node will refuse to receive the coing packet. This eans, when the network is in the congestion state, this policy will at least reduce half of the network traffic. This is designed for ulti-copy routing schees. Because there are ultiple copies of packets in the network, reducing half of the network traffic will not reduce the delivery ratio. ontrarily, it will help iproving the delivery ratio when in congestion state. IV. SIMULATION RESULTS AND ANALYSIS The ONE siulator is used for our siulation. We eploy Epideic Routing protocol to do the siulation. In the results of the siulation, we focus on delivery ratio, average latency, average hop count and overhead ratio. The delivery ratio and overhead ratio are defined as follows: delivery_ratio = delivered / created (6) overhead_ratio = (relayed-delivered)/delivered (7) The delivered is the nuber of packets received by destination nodes. The created is the nuber of packets created by source nodes. The relayed is the total nuber of packets relayed in the network. The average latency is average End-to-End latency of delivered packets, and the average hop count is the average hop count of delivered packets too. There are two scenarios in our siulation. The first scenario is the Ad Hoc scenario. There is only one group of nodes in the Ad Hoc scenario. The settings of every node are the sae. The second scenario is the city scenario. This scenario siulates the Helsinki city. There are six groups of nodes in the scenario. Different settings are deployed for different groups. In the following experients, we set =.5, =.5, α =.5, and r t =.8. A. The Ad Hoc Scenario In this scenario, we assue that all the nodes are with the sae counication ranges, and when every two nodes are in counication range, they can setup the connection, and stop transferring when they are out of counication range. There are ten rounds in the siulation. The node buffer size is MB for the first round and MB for the second round, and so on. The concrete paraeter settings are list in table I. Transission range Message size Message TTL Routing protocol Scenario size 5 Nuber of nodes 5 Siulation tie Message creation interval ~5 KB 6in Epideic Routing h One new essage every 8 to s We copare the strategy with the ost used strategy. The experient results in the Ad Hoc scenario are shown as follows: Delivery ratio Figure. Delivery ratio perforance in the Ad Hoc scenario Fro Figure, we can find that our strategy significantly iproves the delivery ratio. opared to the strategy, the average iproveent of the delivery ratio of the strategy is around 77%. When the buffer size is 4MB, copared to the strategy, the best iproveent of the delivery ratio of the strategy is about 94%. These results are ainly because is based on better buffer anageent policy and better packet receiving policy. TABLE I. SETTINGS OF THE AD HO SENARIO Node speed Transission speed.5~.5/s 5KB/s

4 Proceedings of the nd International onference on oputer Science and Electronics Engineering (ISEE 3) Average latency (s) Figure. Average latency perforance in the Ad Hoc scenario Fro Figure, we can find that the average latency perforance of the strategy is siilar to the strategy. The average latency perforance of the strategy is not good enough, because the latency estiation value in forula () is not stable. When the siulation tie increases, even if there is no transission in the network for a while, the latency estiation value in forula () increases too, so the average latency perforance of the strategy is not very well. Average hop count Figure 3. Average hop count perforance in the Ad Hoc scenario Fro Figure 3, we can find that the average hop count perforance of the strategy is better than the strategy. This is because the strategy drops packets with ore hops and larger latency first. The hop count estiation value in forula () is unlike the latency estiation value. If there is no transission in the network for a while, the hop count estiation value will not change. Overhead ratio Figure 4. Overhead ratio perforance in the Ad Hoc scenario Fro Figure 4, we can find that the strategy gains lesser overhead ratio, copared to the strategy. The overhead ratio of the strategy is around, and the overhead ratio of the strategy is around. The overhead ratio perforance of the strategy is very good, because the packet receiving policy of the strategy reduces ost useless transission in the network. The strategy akes nodes just receive eaningful packets, which are with less hops and lower latency. These experient results in the Ad Hoc scenario show that: the strategy significantly iproves the delivery ratio and reduces the overhead ratio, copared to the strategy. The average hop count perforance of is good, but the average latency perforance of is not very well. B. The ity Scenario The city scenario is totally the default scenario in the ONE siulator. The city scenario is ore coplex than the Ad Hoc scenario. The city scenario siulates the Helsinki city. There are six groups of nodes in the network. Pedestrians (group and group 3) can ove anywhere in the ap, but cars (group ) can only ove on roads. There are three groups of tras in the city scenario, group 4 is with broadcast interfaces, and group 5 and group 6 are noral. Most of the nodes in the network are pedestrians and cars. There are ten rounds in the siulation. For group, group and group 3, the node buffer size is MB for the first round and MB for the second round, and so on. For the groups of tras, the node buffer size is always 5MB. The concrete paraeter settings are list in table II. Node speed TABLE II. SETTINGS OF THE ITY SENARIO Group :.5~.5/s, 4 pedestrians Group :.7 /s ~3.9 /s, 4 cars only on roads Group 3:.5 /s ~.5/s, 4 pedestrians Group 4: 7 /s ~ /s, tras with broadcast interfaces 3

5 Proceedings of the nd International onference on oputer Science and Electronics Engineering (ISEE 3) Group 5: 7 /s ~ /s, tras Group 6: 7 /s ~ /s, tras Transission speed 5KB/s (MB/s for group 4) Transission range ( for group 4) Message size Message TTL Routing protocol Scenario size Nuber of nodes 6 Siulation tie Message creation interval 5KB ~ MB 3in Epideic Routing h One new essage every 5 to 35s We copare the strategy with the ost used strategy. The experient results in the city scenario are shown as follows: Delivery ratio Average latency (s) Figure 6. Average latency perforance in the city scenario Fro Figure 6, we can find that the average latency perforance of the strategy is siilar to the strategy. The average latency perforance of the strategy is not good enough, because the latency estiation value in forula () is not stable. When the siulation tie increases, even if there is no transission in the network for a while, the latency estiation value in forula () increases too, so the average latency perforance of the strategy is not very well. 6.. Figure 5. Delivery ratio perforance in the city scenario Fro Figure 5, we can find that, when the buffer size becoes larger, our strategy significantly iproves the delivery ratio. opared to the strategy, the average iproveent of the delivery ratio of the strategy is around 86%. When the buffer size is 5MB, copared to the strategy, the best iproveent of the delivery ratio of the strategy is about 3%. These results are ainly because is based on better buffer anageent policy and better packet receiving policy. When the buffer size is MB, the iproveent is not obvious, because there are so few packets (usually one or two) in the buffer to use for the latency estiation and the hop count estiation. When packets in the buffer are too few, the strategy will not work well. Average hop count Figure 7. Average hop count perforance in the city scenario Fro Figure 7, we can find that the average hop count perforance of the strategy is better than the strategy. This is because the strategy drops packets with ore hops and larger latency first. The hop count estiation value in forula () is unlike the latency estiation value. If there is no transission in the network for a while, the hop count estiation value will not change. 4

6 Proceedings of the nd International onference on oputer Science and Electronics Engineering (ISEE 3) Overhead ratio Figure 8. Overhead ratio perforance in the city scenario Fro Figure 8, we can find that the strategy gains lesser overhead ratio, copared to the strategy. When the buffer size is MB, the strategy reduces about half of the overhead ratio. Because the packet receiving policy at least reduces half of the network traffic, even if the iproveent of delivery ratio is not obvious, still significantly reduces the overhead ratio. The overhead ratio perforance of the strategy is very good, because the packet receiving policy of the strategy reduces ost useless transission in the network. The strategy akes nodes just receive eaningful packets, which are with less hops and lower latency. These experient results in the city scenario show that: the strategy significantly iproves the delivery ratio and reduces the overhead ratio, copared to the strategy. But when there are too few packets in the buffer, the iproveent of delivery ratio of is not obvious. The average hop count perforance of is good, but the average latency perforance of is not very well. [4] A. Vahdat and D. Becker, Epideic routing for partially connected ad hoc networks, Technical Repor t S-6, Tech. Rep.,. [5] The Opportunistic Network Environent siulator hoepage [.6/.6], [6] E. oe,. Raghavendra, Token Based ongestion ontrol for DTNs, IEEE Aerospace onference, pp. 7, [7] Daowen Hua, Xuehui Du, Lifeng ao, Guoyu Xu, Yanbin Qian, A DTN congestion avoidance strategy based on path avoidance, nd International onference on Future oputer and ounication (IF), vol., pp , [8] F. De Rango, M. Tropea, G.B. Laratta, S. Marano, Hop-by-Hop Local Flow ontrol over InterPlaNetary Networks based on DTN Architecture, IEEE International onference on ounications, I '8, pp. 9 94, 8. [9] A. Krifa,. Barakat, T. Spyropoulos, An optial joint scheduling and drop policy for Delay Tolerant Networks, World of Wireless, Mobile and Multiedia Networks, pp: 6, 8. [] L. Yun,. Xinjian, L. Qilie, Y. Xiaohu, A Novel ongestion ontrol Strategy in Delay Tolerant Networks, Second International onference on Future Networks, IFN ', pp ,. [] P. Gupta, P.R. Kuar, The capacity of wireless networks, IEEE Transactions on Inforation Theory, pp , V. ONLUSION Delay Tolerant Networks often use ulti-copy routing schees for essage transission. Multi-copy routing schees, such as Epideic Routing, usually lead to congestion probles. In this paper, we present a new congestion avoidance strategy to solve congestion probles. The siulation results show that the strategy significantly iproves the delivery ratio and reduces the network overhead. The average hop count perforance of is good too. It is our future work to iprove the strategy for reducing the network latency. REFERENES [] K. Fall, A delay tolerant network architecture for challenged internets, in Proc. AM SIGOMM, pp.7-4, 3. [] DTN research group hoepage[.6/.6], [3] K. Scott, S. Burleigh, RF55: Bundle Protocol Specification, Noveber 7 5