H-ABC: A scalable dynamic routing algorithm

Transcription

1 Chapter 1 H-ABC: A scalable dynamic routing algorithm B. Tatomir and L.J.M. Rothkrantz COMBINED Project, DECIS Lab, Delftechpark 24, 2628 XH Delft, The Netherlands, b.tatomir@ewi.tudelft.nl Man-Machine Interaction, Faculty of Computer Science and Electrical Engineering, Delft University of Technology, Mekelweg 4, 2628 CD Delft, The Netherlands l.j.m.rothkrantz@ewi.tudelft.nl In small networks, ant based algorithms proved to perform better than the conventional routing algorithms. Their performance decreases by increasing the number of nodes in the network. The scalability of the algorithms is affected by the increasing number of agents used. In this paper we present a scalable Hierarchical Ant Based Control algorithm (H-ABC) for dynamic routing. The network is split into several smaller and less complex networks called sectors. The agents are divided in two categories: local ants and exploring ants. Only the nodes situated at the border between sectors can generate exploring ants, the ones used to maintain the paths between different sectors. They are carrying no stack which reduces the overhead in the network. The algorithm was implemented and its performance compared with the well known AntNet. 1.1 Introduction The AntN et adaptive agent-based routing algorithm [3; 4], is the bestknown routing algorithm for packet-switched communications networks, 1

2 2 The Second Australian Conference on Artificial Life 2005 which is inspired from the ants life. Besides the probability tables, at each node the average trip time, the best trip time, and the variance of the trip times for each destination are saved. Routing is determined through complex interactions of network exploration agents. These agents (ants) are divided into two classes, the forward ants and the backward ants. The idea behind this sub-division of agents is to allow the backward ants to utilize the useful information gathered by the forward ants on their trip from source to destination. Based on this principle, no node routing updates are performed by the forward ants, whose only purpose in life is to report network delay conditions to the backward ants. This information appears in the form of trip times between each network node. The backward ants inherit this raw data and use it to update the routing tables of the nodes. In [1] AntN et was improved with an intelligent routing table initialization, a restriction on the number of ants in the network and a special pheromone update after node failures. An increased adaptivity of ants [11] and reduced size of the routing tables [7] was achieved by combining AntN et with genetic algorithms. Ant-Based Control (ABC) was the first successful swarm based routing algorithm and designed for telephone networks [8]. This algorithm uses only one class of forward ants, which travel along the network gathering information. They use this information for updating the routing table rows corresponding to their source node. In [2] ABC was extended and the pheromone updating was introduced also for the other intermediate nodes visited by the agent. Two other ramifications of ABC for packet-switched networks were presented in [9] and [5]. Although proved to perform better than the best classic algorithms like RIP (Routing Information Protocol) and OSP F (Open Shortest Path First), AntNet and ABC have scalability problems [6]. Since each node has to send an ant to all the other nodes of the network, for large networks, the amount of traffic generated by the ants would be prohibitive. Furthermore, for distant destinations there is a larger likelihood of the ants being lost. Moreover, the large traveling times of the ant render the information they carry outdated. One way to solve this load problem and attain scalability is by using hierarchical routing. Adaptive-SDR [6] groups nodes into clusters and directs data packets from a source node to a destination by using intra and inter-cluster routing. Two types of agents, are introduced into the network. The first type is colony ants and the second type is local ants. The task of the colony ants is to find routes from one cluster to the other, while local ants are confined within a colony and are responsible for finding routes within their colonies. The colony ants are launched at every node. This keeps the overhead high.

3 Hierarchical routing in traffic networks 3 BeeHive [10] is a novel routing algorithm, which has been inspired by the communicative and evaluative methods and procedures of honey bees. In this algorithm, bee agents travel through network regions called foraging zones. The model requires only forward moving agents and they utilize an estimation model to calculate the trip time from their source to a given node. The agents are sent to other nodes by broadcasting. Each node keeps a copy of each agent. Although they are really small the overhead is higher than in case of AntNet. Another real disadvantage is the higher memory use for storing every agent to distinguish between different replicas received at the same node. In the next section we introduce a new hierarchical routing algorithm (H-ABC) which combines features of the algorithms presented before. We tested and compared its performance with AntNet in a network with 114 nodes and 172 bidirectional links. The simulation environment and the experimental settings are presented in section 3. The tests and results are described in section 4. The last section contains conclusions and future work. 1.2 The Hierarchical Ant Based Control algorithm Network model We decided to split the network in sectors. The nodes situated at the border of a sector and which have connection with other sectors are called routing nodes. As we will see later these nodes will play a special role, their activity being different than the one of an inner sector node. An example of such a network is shown in Figure 1.1 representing the Japanese Backbone NTTNet divided in 3 sectors. The routing nodes are circled. Fig. 1.1 Japanese NTTNet

4 4 The Second Australian Conference on Artificial Life 2005 A common feature of all the routing algorithms is the presence in every network node of data structure, called routing table, holding all the information used by the algorithm to make the local forwarding decisions. The routing table is both a local database and a local model of the global network status. Each node i in the network has a probability table for every possible final destination d. The tables have entries for each next neighbouring node n, P dn. This expresses the goodness of choosing node n as its next node from the current node i if the packet has to go to the destination node d. In our case this data structure has to be modified. For every sector of the network a virtual node is introduced. This can be understood as an abstraction for all the nodes of the sector. Each virtual node will have an entry in the data structures of every node. They will be used to route the data between different sectors. Considering the network with 3 sectors in Figure 1.1, an example of a routing table in node 12 of a sector will look like this. Table 1.1 Routing table in node 12 Neighbour Destination N10 N11 N13 N15 D D D D D V V Besides the probability table in every node i we have the following additional data structure: µ d : an array storing the mean value of the delay encountered for the destination d S[d]: an array which maps every node in the network to the corresponding sector. U[d]: an array of flags which mention if a node d is up or down The algorithm makes use of 3 types of ants: local, backward and exploring ants. Next we will describe the behaviour of each entity present in the network.

5 Hierarchical routing in traffic networks Local ants The purpose of the local ants is to maintain the routes between nodes of the same sector or to the closest routing node for another sector. Each node s inside a sector periodically generates a local ant F sd. The destination d can be another node in the sector or a virtual node: a node in the same sector(s[d] = S[s]) with probability of 0.9 d = a virtual node(s[d] S[s]) with probability of 0.1 (1.1) The routing nodes have a different behaviour. In this case: local ant(s[d] = S[s]) with probability of 0.1 F sd = exploring ant (S[d] S[s]) with probability of 0.9 (1.2) A local ant behaves similar with the forward ant in AntNet. It keeps a memory about the visited nodes and the estimated time to reach them. At each node i, before going to the next neighbour n, the ant memorizes the identifier of the next node n and a virtual delay T in. This delay represents the estimative time necessary for the ant to travel from node i to n using the same queues as the data packets. T in = q n + Size(F sd ) B in + D in (1.3) q n [bits] is the length of the packets buffer queue towards the link connecting node i and its neighbour n; B in is the bandwidth of the link between i and n in [bit/s]; Size(F sd ) [bits] is the size of the local ant; D in is the propagation delay of the link. The selection of the next node n, to move to, is done according with the probabilities P d and the traffic load in the node i. P dn = P dn + α l n, α = 0.4 [0, 1] (1.4) 1 + α ( N i 1) l n [0, 1] is a normalized value proportional to the amount q n (in bits waiting to be sent) on the link connecting the node i with its neighbour n:

6 6 The Second Australian Conference on Artificial Life 2005 l n = 1 q n Ni j=1 q j (1.5) If a cycle is detected, the cycle s nodes are popped from the ant s stack and all the memory about them is destroyed. If the cycle is greater than half the ant s age, the complete ant is destroyed. A local ant is not allowed to leave his sector. In this way in all the probability tables of the routing nodes we have P dn = 0 (see Table N15) if: S[d] = S[i]: the destination is a node inside the sector S[n] S[i]: the neighbour n is in another sector For a local ant there are two possibilities to reach its destination. One is of course when it arrives in the node d. But when d is a virtual node it stops at the first encountered routing node. In this case it pushes on the stack the node d identifier and µ d, the average time to go from node i to the sector d. At this moment the agent F sd finishes its trip. It transfers all of its memory to a new generated backward ant B ds and dies Backward ants A backward ant takes the same path as that of its corresponding local ant, but in the opposite direction. At each node i along the path it pops its stack to know the next hop node. It updates the routing tables for the node d but also for all the subpaths from i d. The time T id to reach d is the sum of all the segments T jk on the path: First it modifies the value of µ d. T id = T jk, where j, k i d (1.6) µ d = µ d + η(t id µ d ), η = 0.1 (1.7) After this it updates the probability table with a reinforcement value r. This is a function of the time T id and its mean value µ d. T id cµ d ; T id cµ d < 1, c = 1.1 > 1 r = r = 1 otherwhise (1.8)

7 Hierarchical routing in traffic networks 7 P dn = P dn + (1 r)(1 P dn ) for n the node chosen by the ant (1.9) P dj = P dj (1 r)p dj, for j n (1.10) When the source node s is reached again, the agent B ds dies Exploring ants The purpose of exploring ants is to find and maintain the routes between different sectors. They are light and keep no track about the path they followed. The only information they register is an estimate time to reach their source sector T s. The exploring ants are generated only by the routing nodes of each sector s. They receive as destination d a virtual node representing another sector. In this case we will refer by s not to the source node but to the sector. They are forwarded to destination using the same mechanism as the local ants. If a cycle is detected it is removed, and in case it is bigger than half of the ant age, the ant is killed. As the local ants which are not allowed to get out of the current sector, the exploring ants are not allowed to move to a node inside their source sector. Once they left the home sector they can t return. If this still happens the exploring ant is killed. This is because there are other routing nodes in the same sector which are closer to the destination sector d of the ant. The ants generated there will be more efficient with that destination. When an exploring ant arrives in a node i coming from a node p, it adds to its trip time T s the trip time necessary for ant to travel from i to p. T s = T s + T ip (1.11) T ip = q p + Size(F sd ) B ip + D ip (1.12) The new computed T s value is used to update the routing table at node i. The changes are are similar with the ones made by the backward ants, but they are done only for the source sector s. µ s = µ s + η(t is µ s ) (1.13)

8 8 The Second Australian Conference on Artificial Life 2005 T s T cµ s ; s cµ s < 1 r = r = 1 otherwhise (1.14) The reinforcement is given to the link i p: P sp = P sp + (1 r)(1 P sp ) (1.15) P sj = P sj (1 r)p sj, for j p (1.16) An exploring ant ends its trip when arrives at a routing node of the destination sector d Data packets A packet P sd generated at a node s can have as destination d any node in the network. Arriving at a node i they are routed according with the probabilities P d. If the destination node d can not be found in the same sector with the node i then the packet is routed to the destination sector S[d] using the correspondent virtual node entry in the routing table Flag packets When a neighbouring node n is detected as inactive, a special packet Y n is broadcasted in the network to inform the other nodes. It is carrying the identifier of the node n and a flag set to down in this case. When the crashed node becomes active again it broadcast the same type of packet but with the flag set on up. When a packet Y n is received at a node i it compare the U n value with his flag value: if U n = Y n, then the change was already done by another packet and the current packet is destroyed. if U n Y n, then set U n = Y n and send Y n to all the neighbours using the same priority queues as the ants. Next we present the pseudocode of the algorithm Algorithm 1.1 H-ABC {i - current node, d - destination node, s - source node}

9 Hierarchical routing in traffic networks 9 {n - successor node of i, p - predecessor node of i} for all N odes {concurrent activity over the network} do if time to generate an agent at node s then Create(F sd ) for all forward ants F sd received at node i do if big cycle detected then kill F sd if F sd is local then dest reached false if d is a virtual sector and i is a routing node then F sd (d, µ d ) {add the virtual node on the stack} dest reached true if d = i then dest reached true if dest reached then create B is send B is to p using the priority queue kill F sd n GetNext(F sd ) T in GetV irtualtime(n, q n ) {compute the virtual time to get to n} F sd (n, T in ) {add the new information on the stack} send F sd to n using priority queue T ip GetV irtualtime(p, q p ) T is T ps + T ip UpdateP robabilities(i, p, T is ) if i is routing node and S[i] = d {destination was reached} then kill F sd n GetNext(F sd ) if S[n] = S[s] and S[i] S[s] then kill F sd {a forward agent should not return to his source sector}

10 10 The Second Australian Conference on Artificial Life 2005 send F sd to n using priority queue end for for all backward ants B sd received at node i do UpdateP robabilities(i, d, T id ) {do updates to the node information} if d i {destination not reached} then n GetNext(B sd ) select next node to go send B sd to n using high priority queues kill B sd end for for all data packet P sd received at node i do if d i {destination not reached} then if S[i] = S[d] {the destination d is in the current sector} then n GetNext(P sd ) n GetNext(P ss[d] ) send P sd to n using normal queue end for for all flag packet Y n received at node i do if U n = Y n then kill Y n U n Y n broadcast Y n using high priority queues end for end for 1.3 Simulation environment We implemented and tested our new H-ABC algorithm in a new developed simulation environment. For comparison we choose the AntN et algorithm which we already had implemented. The common parameters of the algo-

11 Hierarchical routing in traffic networks 11 rithms were set to the same values as for AntNet in [4]. In the literature, the most complex network instance that was mostly used in simulations is the Japanese Internet Backbone (NTTNET)(see Figure 1.1). It is a 57 node, 81 bidirectional links network. The link bandwidth is 6 Mbits/sec and propagation delay rages from 0.3 to 12 milliseconds. In order to create a more challenging environment we linked two copies of the NTTNet network and added 10 extra links (see 1.2). The result is a network with 114 nodes and 172 bidirectional links. Traffic is defined in terms of open sessions between two different nodes. At each node the traffic destination is randomly chosen between the active nodes in the network and remains fixed until a certain number of packets (50) have been sent in that direction. The mean size of data packet is 4 Kbytes, the size of an agent is 192 bytes. The queue size is limited to 1 Gb in all experiments. We studied the H-ABC performance related to the number of sectors the network was divided. 3 versions of the H-ABC algorithm were tested: H-ABC2: network with 2 sectors H-ABC4: network with 4 sectors H-ABC6: network with 6 sectors Actually the version of the algorithm for only one network (H-ABC1) is exactly AntN et. Fig. 1.2 Traffic network with three cities

12 12 The Second Australian Conference on Artificial Life Test and results During the simulations, we focused on 4 metrics for evaluation: the troughput, the delivered packets rate, the arrived packets delay and the overhead. Each test lasted 300s with a training period (just ants were generated) of 30s Low traffic load A routing algorithm should perform well not only under heavy traffic load but also under low load over the network. To achieve such a traffic we set the mean packet generation period(pgp) to 0.3s. Figure 1.3 shows the average packet delay. The average packet delay for H-ABC6 and H-ABC4 is below 250ms, around 250ms for H-ABC2 and above 300ms for AntNet. All the algorithms delivered similar throughput and more than 99% of the packets. Fig. 1.3 Average packet delay under low traffic load High traffic load For this test we decreased the mean packet generation period to 2ms. Again H-ABC4 and H-ABC6 scored the best. They delivered 99% of the packets with an average delay below 1s. H-ABC2 delivered 98% of the packets and after an increasing slope the average delay went down to 1s. The AntNet had not so good performance. Just 83% of the packets reached the destination with an average delay of more than 5s. As expected from the difference between the packets delivered ratios,

13 Hierarchical routing in traffic networks 13 Fig. 1.4 Average packet delay under high traffic load the throughput of H-ABC algorithms 2250Kb/s 1 is higher than the average throughput of AntNet 1900Kb/s 1. Fig. 1.5 Average throughput under high traffic load Hot spot We started the simulation with PGP=100ms. After 100s the node 54 became a hot spot : all the nodes started to send the packets to it with PGP=75ms. The hot spot was active for 100s. All 4 algorithms behaved well under this test delivering 99% of the packets. The only difference was again the average packet delay (Figure 1.6).

14 14 The Second Australian Conference on Artificial Life 2005 Fig. 1.6 Average packet delay under hot spot state Overhead Finally we measured the overhead of the algorithms. For this we run a short 50s test with only agents running in the network and measured bandwidth capacity usage. The overhead decrease with the number of sectors. AntNet agents used about 0.375%, H-ABC %, H-ABC % and H-ABC % of the network capacity. 1.5 Conclusions and future work We combined features from several ant based algorithms and obtained a highly scalable and robust algorithm. Its performance was tested in a simulation environment and compared with AntN et. H-ABC performed better both in low and high traffic load, but also in case of transient overloads over the network. With a low overhead it delivered faster the packets to destination, decreasing also the number of lost ones. Dividing the network in sectors was very effective. The results of the H-ABC4 and H-ABC6 are similar so a big fragmentation of the network is not necessary. Our research is focused on developing a Routing System for cars in traffic networks. At TUDelft we have a city traffic environment which uses a version of AntN et for routing the vehicles. The scalability problem because of the big size of the city maps made us to look for a better scalable algorithm like the one presented in this paper. We expect it to be effective in our case because of the hierarchical distribution already existent in the current traffic network (cities linked by motorways).

15 Bibliography B. Baran. Improved antnet routing. SIGCOMM Comput. Commun. Rev., 31(2):42 48, E. Bonabeau, F. Henaux, S. Guerin, D. Snyers, P. Kuntz, and G. Theraulaz. Routing in telecommunications networks with ant-like agents. In IATA 98: Proceedings of the second international workshop on Intelligent agents for telecommunication applications, pages 60 71, G. Di Caro and M. Dorigo. Antnet: A mobile agents approach to adaptive routing. Technical Report IRIDIA, Universit Libre de Bruxelles, (12), G. Di Caro and M. Dorigo. Antnet: distributed stigmergetic control for communication networks. Journal of Artificial Intelligence Research (JAIR), 9: , M. Heusse, D. Snyers, S. Gurin, and P. Kuntz. Adaptive agent-driven routing and load balancing in communication networks. ENST de Bretagne Tech. Rep, I. Kassabalidis, M.A. El-Sharkawi, R.J. Marks II, P. Arabshahi, and A.A. Gray. Adaptive-sdr: Adaptive swarm-based distributed routing. In World Congress on Computational Intelligence, S. Liang, A. Zincir-Heywood, and M. Heywood. Intelligent packets for dynamic network routing using distributed genetic algorithm. In Genetic and Evolutionary Computation Conference, GECCO, R. Schoonderwoerd, O.E. Holland, J.L. Bruten, and L.J.M. Rothkrantz. Antbased load balancing in telecommunications networks. Adaptive Behavior, (2): , D. Subramanian, P. Druschel, and J. Chen. Ants and reinforcement learning: A case study in routing in dynamic networks. In The Fifteenth International Joint Conference on Artificial Intelligence, pages , H. Wedde, M. Farooq, and Y. Zhang. Beehive: An efficient fault-tolerant routing algorithm inspired by honey bee behavior. In 4th International Workshop, ANTS 2004, T. White, B. Pagurek, and F. Oppacher. ASGA: Improving the ant system by integration with genetic algorithms. In Genetic Programming 1998: Proceedings of the Third Annual Conference, pages ,