A Traffic Predictor Based on Adaptive Fuzzy Clustering Technique for IP-based Networks

全文

(1)A Traffic Predictor Based on Adaptive Fuzzy Clustering Technique for IP-based Networks Yau-Hwang Kuo and Mong-Fong Horng* Department of Computer Science and Information Engineering, National Cheng Kung University. Department of Electrical Engineering, Kao-Yuan Institute of Technology* Email : [email protected]. Abstract. the Internet, we need a mechanism consisting of admis-. A traffic predictor is significant to ensure the reliability. sion control, flow control and resource management to. and utilization of network resource in the next generation. guarantee the end-to-end Quality of Service (QoS) for. Internet. An adaptive approach based on fuzzy clustering. users is urgently needed. Unfortunately, the existed. technique for traffic estimation in resource protocols is. IP-based Internet lacks some significant properties to. proposed. The proposed traffic estimator is deployed to. realize the goal of offering feasible multimedia services. learn the knowledge about the traffic flows and generates. with tolerable packet delay and loss. To solve this prob-. a set of cluster representing the profile of flow character-. lem, many new architectures and protocols for resource. istics. Different from the past k-mean clustering, the pro-. reservation had been presented. In the reservation proto-. posed clustering algorithm does not need the prior infor-. col, discriminative delivery is the most important mecha-. mation of cluster number. The size of cluster set is fully. nism for developing QoS delivery. All flows should be. determined by target pattern set. By matching the derived. admitted to have a reasonable sharing of network band-. cluster set of flow patterns with the current flow pattern,. width before starting transmission. Obviously, how to. the future flow is estimated and predicted. In experiments,. determine the bandwidth sharing is the key point to QoS. two kinds of traffic flow are used to verify the algorithm. guarantee. The conventional scheduler serving in a steady. performance, periodic flow and Poisson flow. Results. round-robin is incapable of dynamic bandwidth allocation.. show excellent performance of the proposed predictor.. In this paper, we propose a traffic predictor based on a. The prediction errors, in average, are near 15.4% and. novel fuzzy clustering technique to assist the traffic. 3.56% for periodic flow and Poisson flow, respectively.. scheduler of reservation protocol, such as RSVP. In the new generation Internet, network resources are no longer. I.. INTRODUCTION. to be used without permission and arbitrary allocation. All network resources, especially the bandwidth, have to. Interactive multimedia communication has become. be allocated by a resource administrator for all connection. one of the most promising technologies on the Internet.. requests to guarantee the admitted connection to obtain an. To provide good multimedia communication services on. agreed service quality. In such a new networking archi-. * The co-author also is a PhD. candidate of the department of computer science and information engineering at National Cheng-Kung University.. tecture, traffic predictor is frequently introduced to help the scheduler decide the transmission order of all incom-.

(2) ing effectively and efficiently so that the congestion can. ity-of-Service (QoS) and bandwidth requirements. In the. be avoided [1]. Consequently, each admitted connection. discussion of QoS issues, datagram loss probability and. could be served with the beforehand promised QoS level.. end-to-end delay are concerned most. For instance, one of. Because of the traffic variation, the bandwidth of the. the most common-used video compression standards is. networking device for each connection also has to be al-. the MPEG standard defined by ISO Motion Picture Ex-. located dynamically to meet the promised QoS level. Be-. pert Group. In a MPEG video service connection, the. sides, since the traffic condition in a connection depends. traffic rate is viewed as a variable bit rate (VBR) traffic. on different applications, the data flow rate of an admitted. with high peak rates relative to its average rate. To main-. connection varies with time in multimedia applications,. tain a stable frame rate at replay, the end-to-end delay of. such as VoD and I-phone. The varying flow rate results in. all datagram should be guaranteed and the datagram de-. the difficulties of traffic scheduling, buffer control and. layed out of bound will be discarded. Thus a sufficient. data loss handling.. bandwidth is required to solve the delay problem. How-. In the previous studies, most of the traffic prediction. ever, excessive bandwidth may cause low bandwidth. focused on multimedia application on ATM [2,3] by us-. utilization. Some efforts tried to smooth a VBR traffic. ing neural network [2,5-6] or fuzzy logic [1,3,7] respec-. stream to a constant bit rate (CBR) traffic, but they often. tively. However, the applicability of neural network ap-. brought on redundancy in realization and caused data-. proach is argued [8] because a neural network has to learn. gram loss while the flow rate is in a bursty. In multimedia. a complicated mapping between past and future arrivals.. service, the diverse characteristic of multimedia bitstream. In the proposed approach, the traffic patterns are pre-. makes it difficult to offer an efficient and effective allo-. processed by the proposed fuzzy clustering to extract the. cation of bandwidth for all admitted connections. The. traffic characteristics. By using the proposed fuzzy clus-. proposed predictive scheduler is shown in Fig. 1. Differ-. tering technique, the representative clusters of traffic pat-. ent from non-predictive schedulers, we use a traffic esti-. terns are obtained to indicate the characteristics of the. mator to predict the rate of flows by analyzing the arrival. flow traffic. Based on the derived cluster set, arrived flow. patterns for each connection. The traffic estimator based. pattern is used to estimate traffic flow and predictor the. on a novel fuzzy clustering technique extracts knowledge. future flow rate precisely. By the proposed clustering. about the target flow characteristic. By using the learned. approach, the proposed traffic estimator features high. knowledge, a pattern-matching scheme is deployed to. estimation speed, high estimation accuracy and high. rapidly predict the future flow rate.. adaptability, so it is suitable for various interactive multimedia communication services on high-speed networks.. A new architecture of QoS router for the next generation Internet is shown in Fig. 1. There are three main. 2. ADAPTVE TRAFFIC PREDICTOR BASED ON FUZZY CLUSTERING. components to be realized. They are a packet classifier, traffic estimator and packet scheduler. An adaptive clustering algorithm is proposed to develop high-performance. High-Speed network supporting multimedia services. traffic predictor in this paper. By mean of the robust. are able to handle bursty traffic and satisfy various Qual-. nonlinear decision of fuzzy logic, it helps to evaluate.

(3) rapidly the flow rate, promote the bandwidth utilization. µ ij =. and improve packet loss. Consider an arbitrary traffic. 1 1 + ci − c j. (3). flow f(t) to be estimated and predicted, we have a sam-. where µij is a menbership grade indicating the. pled flow value f(k△t), for k=0, 1, 2,…and △t is the. similariry between c1, and c2 as well as | . | is an. period of sampling flow pattern. For simplicity, f(k△t) is. Euclidean distance between two patterns. denoted as f(k). From the flow sample f(k), a set of flow Step 2. Given a flow pattern pj, we refresh the evaluation. patterns P={p(k)} is obtained by. matrix by P(k)={f(k△t), f((k-1) △t)}, …, f((k-n+1)△t ). M K E=  U. (1). In different kinds of Internet service, diverse characteris-. UT   1 . and. tic is observed and analyzed generally in statistics to de-. U = (u K +1,1, u K +1,2 ,..., u K +1, K ). rive either mathematical correlation or flow tendency.. (5). The complicated and inflexible model usually result is an infeasible solution.. (4). where u ij =. 1 1 + ci − p j. is the silimarity grade. In our approach, an adaptive clustering algorithm is. representing the slimiarity between ci and pj, UT is. developed to derive the correlation between successive. the transport vector of U and N is the dimension. flow patterns and to realize the traffic prediction. Clus-. of square matrix M. It is obvious that E is an unity. tering algorithm is used to cluster the generated traffic. diagonal square matrix with dimension (K+1).. patterns and to find the desired characteristic. Fuzzy c-means is a well-known clustering algorithm to analyze. Step 3. Evaluate the possibility of the new patttern. a given data set. However, we have to specify the cluster. becoming a cluster center from the element {eij}. number before starting the c-mean algorithm. In network. of matrix E;. applications, the prior knowledge is not available. On. if eK +1, j < min{eij }. behalf of the insufficiency, in the traffic estimation prob-. i≠ j. for j=1, 2, …, K. (5). lem, the flow variation may vary greatly that causes the. then pj will be a new cluster center ck+1 , K is in-. unpredictability about the number of clusters. Thus, we. creased by one and the proximity matrix. propose a new clustering algorithm to release the con-. MK+1 is upgraded according to MK+1 = E. straint on cluster number c as follows. (6). else pj is assigned to cluster i given by Step 1. Initialize two cluster centers c1 and c2 in the pat-. i = max {ei , K +1} 1≤ i ≤ N. tern space as landmarks and generate a proximity matrix. (7). Step. 4 Replace all cluster centers with the cluster. M  1 M2 =  µ 21. and. µ 12   1 . (2). centroid, given by.

(4) The row vector ci is a center derived from the clustering. ni. i Σ pj. ci =. j =1. of flow patterns {p(k)}. A clusteri matching scheme is. (8). ni. used to estimate the traffic flow when a new flow pattern. Step. 5 Goto Step 2.. arrives. To obtain a N-step prediction We partition the matrix C as. The. initial. landmarks. specify. the. minimum. dissimilarity betwween two different cluster centers. The similarity menbership function used here is bounded in the intrval of [0, 1]. In this clustering scheme, the similarity instead of specifying a cluster number in the method of fuzzy c-means, we define a minimum vigilance via the initial landmark in initialization. When a pattern with dissimilarity exceeding the vigilance, a new cluster.  c11  c 2,1 C =   c K ,1  vi  1 =  i  v K. c1,2 c 2, 2. c1,n − N. c1,n − N +1 . c K ,n − N. c K ,n − N +1. c1,n      c K ,n  . v1o     v ko  . (11). is created for this pattern, otherwise this pattern will be classified to the cluster that is the nearest one to it. Meanwhile, the cluster receiving new clasified patterns. The sets of row vectors {vi} and {vo} are the input pattern. will upgrade its center by the new centroid of all patterns. and output patterns of the traffic predictor respectively.. in it. The obtained cluster set is representative for the. That is the traffic predictor will decide the mapping. traffic characteristic of flow f(k) through the clustering. between {vi} and {vo} by matching {vi} with the corre-. process of flow patterns of f(k). Sicen all flow patterns. sponding values of each cluster center. The {vo} derived. would be classified to a certain cluster, the averagement. from the corresponding values of the best-matched cluster. of all flow patterns of a cluster is called cluster center.. center will yield the proper prediction.. Thus the cluster centers means a suitable and reasonable. Because of the high demand on computation efficiency. prediction baseline to find the prediction fk+1 such as the. on network application, the time consumption of traffic. Prob(f((k+1)= fk+1 | f((k)= fk , f((k-1)= fk-1,…, f((k-n+1)=. predictor is usually argued. In this paper, the network. fk-n+1 ) is maximized, where Prob(|) is conditional. training does not occur at each time instance when a pat-. probability.. a. tern is received. The occurrence of clustering process. new-generated flow pattern is matched with the learned. depends on the difference between the receiving flow. cluster centers individually and sequentially to predict the. pattern and the extracted traffic characteristic. When the. future flow rate.. Given a center matrix C with K. difference exceeds the vigilance, the clustering process is. clusters derived from the fuzzy clustering mentioned. activated. In the following experiment, it shows that, in. above,. many traffic conditions, the clustering process is activated. In. the. proposed.  c1     c11 c  C= 2=     c K1 c K . traffic. predictor,. at the short beginning of flow. That is, the proposed trafc12 cK 2. . c1n    c Kn . (10). fic predictor is effective and efficient. Thus, after the predictor has extracted the representative knowledge enough, the occurrence of learning will not increases until the.

(5) traffic characteristic changes. Different from the conven-. proposed adaptive fuzzy clustering algorithm. Clearly,. tional approach, the efficiency of estimator is improved.. the increasing of cluster number with time indicates the learning process of the proposed predictor. However,. 3. EXPERIMENTS. after reaching the level of mutual learning, the cluster number remains steady. It means that the later flow. In order to approve the performance of the proposed. variation does not bring any new learning information.. traffic predictor precisely, three flow conditions are. Comparing Fig.2 with Fig.3, it is very interesting to find. adopted in experiments. The first flow type named peri-. that, in periodical flow, the learning time is shorter than. odical flow that equips either a variation period or a cer-. the flow variation period from various experimental. tain self-similarity. Owing to the encoding characteris-. examples. In other words, the proposed clustering algo-. tics of MPEG video, such kind of flow type is quite. rithm exhibits a very positive performance in learning.. normal in video flow. The second is Poisson flow. In a. The prediction result of Poisson flow is shown in. Poisson flow, the interarrival time of two successive. Fig.4. Unlike periodical flow, since Poisson flow does. packets is characterized by a random variable with. not equip with a periodical variation, it is more difficult. Poisson distribution. Generally, in communication net-. to estimate and predict a Poisson flow than a periodical. works, it is well known that flows are Poisson flows. flow. Nevertheless, the experimental result shows an. with different mean values. The third flow type used in. acceptable performance in Fig. 4. Although the cluster. experiments is real peer-to-peer video flow measured in. number largely increases, the prediction error remains. local area network.. about under 15%. In the third experiment, a real video. The prediction performance on periodical flow is. flow sampled from the film ‘Titanic’ is taken as the test. shown in Fig. 2. The solid line depicts the flow to be. data. The prediction result is shown in Fig. 5. Although. predicted and the dash line shows the result predicted by. the burst prediction is not good enough, the average ac-. the proposed predictor. To verify the prediction per-. curacy should be acceptable.. formance on jounce flow, the target flow is synthesized. Experiencing the three testing flow types mentioned. by sine waves with variant frequencies. From Fig.2, the. above, the traffic predictor based on the proposed adap-. target flow periodically jounces between 1 Mbps and 9. tive fuzzy clustering algorithm demonstrates excellent. Mbps. It should be noted that, in all experiments, the. performance on each test. In particular, it shows an ex-. traffic predictor has no information about the target flow,. tremely good and stable performance on periodical flow.. such as mean rate and the others. Practically, the pro-. Furthermore, from different experiments, the proposed. posed predictor operates in a on-line manner by mean of. clustering algorithm can actually learn new flow char-. receiving the contiguous-varying flow and meanwhile. acteristics adaptively and dynamically which works as. estimating and predicting the flow rate in the next time. the traffic predictor’s knowledge base to promote the. instance. In the experiment on periodical flow, the pre-. prediction performance. Besides, in the proposed clus-. diction error rate is about 3.42 % and the derive cluster. tering algorithm, cluster number is not required any. set is composed of 20 clusters. Fig. 3 shows the evolu-. more before clustering analysis. All we have to do is to. tion of the number of the obtained clusters through the. establish the possible minimal distance between two.

(6) nearest clusters, named vigilance. And then, the pro-. posed predictor depicts that the proposed predictor not. posed clustering algorithm will adaptive increase the. only improves the bandwidth efficiency but reduces the. cluster number on its own. This is a great difference. requirement of buffer size. That is the buffer occupancy is. compared with the previous fuzzy c-mean clustering. slight if the predictor works well. It is also seen that in a. algorithm. non-predictive scheduler, the buffer likely overflows. As mentioned above, in the initialization of adaptive. when a bursty duration arrives. However, scheduling with. clustering, two predefined clusters are used as landmark. our traffic predictor, a slight variation of buffer occu-. to analyze flow patterns. The initial landmarks determine. pancy definitely helps packet loss ratio and network re-. the minimum vigilance of similarity between two differ-. source management, even in case of small buffers.. ent clusters. By adjusting the vigilance, we have different traffic characteristic after clustering. The different sets of. 4.CONCLUSION. initial landmarks caused by different vigilance produce various cluster sets. Table 1 shows predictor performance. In this paper, we also propose a new clustering algo-. on Poisson flows under conditions of different extracted. rithm in comparison with fuzzy c-means, having the ex-. cluster sets.. pressive benefit of releasing the constraint that specifies. Table 1. The prediction performance of different vigi-. the number of cluster in prior. Based on new clustering. lance.. algorithm, we also propose the architecture of adaptive predictor in reservation protocol. Since the proposed. No. of resulted 5 10 20 Clusters Estimation 86.5 89.2 89.7 Accuracy % % %. 40. 80. 160. 87.8 %. 91.4 %. 91.1 %. clustering algorithm can adaptively generate necessary cluster to generalize the characteristic of analyzed data patterns. Thus, the dynamic property of our clustering scheme has the benefits that can handle the successive. In fact, a higher vigilance value could derive fewer clus-. data and dynamically adjust the cluster locations to rep-. ters from traffic flow. That is, a cluster could accommo-. resent the flow pattern characteristic. It is an important. date those patterns with higher dissimilarity. Likewise,. improvement enable clustering to be applicable in con-. the set of fewer clusters accelerates the prediction speed. tinuous dynamic system, such as multimedia network.. in pattern matching. However it may degrade the predic-. While comparing with the nearest neighborhood cluster-. tion accuracy.. ing, our scheme has the variable boundary for each clus-. By mean of traffic prediction, traffic scheduler could. ter that modifies the property of fixed boundary in the. allocate bandwidth of flows in advance according to the. nearest neighborhood clustering. In experiments, the pre-. produced prediction so that the bandwidth efficiency is. sented traffic predictor indeed has an excellent perform-. improved. The non-predictive scheduler, because of the. ance on real VBR streams of video service. The accuracy. lack of the adjustability of bandwidth allocation, either. of predictor is acceptable in network applications. In the. has a low bandwidth utilization or high buffer occupancy.. next generation Internet, a discrimination delivery is the. In Fig. 6, the comparison of the buffer occupancy caused. trend of network protocol design. Network resource. by a packet scheduler with non-predictive and the pro-. should be reserved for admitted connections to guarantee.

(7) QoS in transmission. The traffic predictor helps the re-. [5] W. Lobejko, “Traffic prediction by neural network,”. source reserved more efficiently and dynamically.. Proc. of Military Communication Conference, vol. 2, 1996. , pp. 571-575. REFERENCES. [6] H P. Lin et. al., “Neural networks based traffic prediction for cell discarding policy,” Proc. of. Int’l Conf. on. [1] Qiu, N.” A predictive fuzzy logic congestion avoid-. Neural networks, 1997, pp. 2051-2056.. ance scheme,” Proc. of Global Telecommunications Con-. [7] M. F. Scheffer, J. J. P. Beneke, J. S. Kunicki, “Fuzzy. ference, vol. 2, 1997, GLOBECOM’98, pp. 967-971.. modeling and Prediction of Network Traffic Fliuctua-. [2] A. A. Tarraf, I. W. Habib, T. N. Saadawi, and S. A.. tions,” Proc of IEEE COMSIG, 1995, pp. 41-45.. Ahmed, “ATM multimedia traffic prediction using neural. [8]J. Hall and P. Mars , “The limitations of artificial neu-. networks,” Proc. of Global Data networking, 1993, pp.. ral networks for traffic prediction,” Proc. of Computers. 77-84.. and Communications, 1998, pp.8-12,. [3] Qixiang Pang et. al., “Adaptive fuzzy traffic predictor. [9]K Tutschku et. al. “Spatial traffic estimation and char-. and ite applications in ATM networks,” Proc. of In’l. acterization for mobile communication network design,”. Communication Conference, vol. 3, 1998 pp. 1759-1763.. IEEE. Journal on Selected Areas in Communications, vol.. [4] A. Kolarov, A. Atai and J. Hui, “Applications of. 16, Issue 5, June, 1998, pp. 804-811.. Kalman filter in high-speed networks,” Proc. of Global Telecommunications Conference, vol. 1, Globecom, 1994, pp.624-628..

(8) 160. Flow rate(Kbps). 140 120 100 80 60 40 20 553. 507. 461. 415. 369. 323. 277. 231. 185. 93. 139. 47. 1. 0 Time (ms). Fig. 4. The prediction performance of Poisson flow.. 562. 511. 460. 409. 358. 307. 256. 205. 154. 103. 1. 10.00 9.00 8.00 7.00 6.00 5.00 4.00 3.00 2.00 1.00 0.00 52. Flow rate (Mbps). Fig. 1. A new architecture of QoS router in the next generation Internet.. Time (ms). Fig. 2. The prediction performance of periodical flow.. Fig. 5. The prediction performance of real video flow.. 80. Cluster Number. 70 60 50 40 30 20 10 599. 553. 507. 461. 415. 369. 323. 277. 231. 185. 93. 139. 47. 1. 0 Time (ms). Fig. 3. The evolution of learned cluster number in clustering analysis of periodical flow patterns.. Fig. 6. The buffer occupancy of video flows with predictive scheduling and non-predictive scheduling..

(9)