Analyzing Spatiotemporal Characteristics of Education Network Traffic with Flexible Multiscale Entropy

Transcription

1 Vol. 9, No. 5 (2016), pp Analyzing Spatioteporal Characteristics of Education Network Traffic with Flexible Multiscale Entropy Chen Yang, Renjie Zhou *, Jian Wan, Jilin Zhang and Yuyu Yin School of Coputer Science and Technology, Hangzhou Dianzi University, Hangzhou, China. Key Laboratory of Coplex Systes Modeling and Siulation, Ministry of Education, Hangzhou, China. * rjzhou@ hdu.edu.cn Abstract In this paper, we propose an analysis ethod of spatial and teporal characteristics of education network traffic based on flexible ultiscale entropy (FMSE). As an iproved ethod, flexible ultiscale entropy has a significant iproveent in stability and accuracy over ultiscale entropy (MSE). We analyze network traffic in different tie scales, space scales and traffic levels by building network traffic tie-space analysis odel and using flexible ultiscale entropy as a ethod to quantify coplexity of different network traffic subsequences. The results show that there are distinct characteristics in the coplexity of different levels of network traffic. We also find that the existence of a large nuber of sall network traffic flows has a significant influence on the coplexity of network traffic. Keywords: education network, network traffic, flow, ultiscale entropy, flexible ultiscale entropy, spatioteporal scale transforation 1. Introduction With the continuous developent of the Internet, it has becoe an indispensable part of people's life and odern education. How to ensure safe, stable and efficient operation of education network has becoe an urgent proble needs to be solved. As the carrier of the data flow in network, the spatioteporal characteristic of network traffic can help to understand the coplex network structure and the dynaic characteristics of network, and it has iportant significance for designing, ipleenting and onitoring network. The self-siilarity and long range dependence [1, 2], have been a hot research topic since they were discovered. As an iportant atheatical tool, the wavelet transfor has been widely applied in the research of network traffic [3]. With the continuous developent of achine learning, any achine learning ethods have been applied into the analysis and research of network traffic [4-6]. However, there are few researches on the coplexity of network traffic. As the network heterogeneity brings the difference of network behavior. Although the IP syste in the user level unify all kinds of heterogeneous network and doain anageent, but the IP layer only shield the heterogeneity of network, and the difference in the network is still exist. According to the research of network coplexity, we can reveal the intrinsic characteristics of network traffic, which can provide a strong theoretical basis for network traffic prediction and anoaly detection. As a easure of syste coplexity, entropy often iplies the level of whole syste coplexity. Entropy theory can help us to discover the inherent dynaics of network * Renjie Zhou and Chen Yang are co-first authors who contribute to this work equally. Renjie Zhou is the corresponding author. ISSN: IJXX Copyright c 2016 SERSC

2 traffic [7]. However, traditional entropy theory can only describe tie series fro a single scale, which cannot reveal the coplexity of tie series copletely. Multiscale entropy cobined entropy theory and ulti scale idea for the first tie, analyzing tie series fro different scales to discovery the correlation of tie series in different scales. Multiscale entropy has been widely applied in different fields, e.g., Janne Riihijärvi use ultiscale to study characteristics of networks and wireless counication [9]. Multiscale entropy has obvious advantages copared with inforation entropy and self-siilar paraeter H, and ultiscale entropy also has a good effect on the prediction of network traffic. In this paper, we use flexible ultiscale entropy as a characterization tool for tie series coplexity. As an iproved algorith of ultiscale entropy, flexible ultiscale entropy has a great iproveent on the stability and accuracy of coputation, and has less dependence on the length of tie series. The rest of the paper is organized as follows. In Section 2, we introduce flexible ultiscale entropy and the tie-space analysis odel in detail. In Section 3, we apply flexible ultiscale entropy and the analysis odel to real network traffic data. In Section 4, we provide the result of traffic statistics and analysis. Finally, we conclude the paper in Section Methodology 2.1. Flexible Multiscale Entropy Saple entropy (SapEn) is a easureent of tie series coplexity proposed by Richan [10], which can be used to quantify the coplexity of tie series. The calculation of tie series coplexity can provide a theoretical basis for the prediction and detection of tie series. Costa found that saple entropy ethod applied to health and disease research often output contradictory results, and then proposed the ultiscale entropy. And experients were carried out with white noise and 1/f noise, the experiental results show that the ultiscale entropy can accurately reflect the coplexity of tie series. However, the stability and accuracy of the ultiscale entropy decreased significantly when the scale factor becoes larger. We proposed flexible ultiscale entropy, which introduces the flexible coefficient f into coputation,and the ethod of flexible accuulation is used in the calculation of saple entropy at the specific scale. By setting the flexible coefficient f to control the atching degree of teplate, the discriination of different tie series is iproved. At the sae tie, the stability and accuracy of coputation iproved greatly copare to ultiscale entropy. The calculation steps of flexible ultiscale entropy are as follows: Step1: A coarse-graining process is applied to a given tie series { x( i) 1 i N}. This process is based on coposite ultiscale entropy proposed in paper [11]. The transfored sequence is as follow: ( ) ( ) ( ) ( ) Nk1 y k { yk,1, yk,2,..., yk, P}, P,1 k (1) For every scale factor, it will build new tie series. The specific transforation forula is as follow: j k1 ( ) 1 yk, j xi,1 j P,1 k (2) i( j1) k Step2: For a given pattern length, coposing each new tie series to diensional consecutive vector sequences, the result is as follow: ( ) ( ) ( ) ( ) Y [ y, y,..., y ],1 i P 1,1 k (3) k, k, i k, i1 k, i1 304 Copyright c 2016 SERSC

3 Step3: We define vector distance d[ Y ( i), Y ( j)] as the axiu value of the ( ) ( ) k, k, difference between corresponding eleents in the two vectors. ( ) ( ) ( ) ( ) d[ Y ( i), Y ( j)] ax y y (4) k, k, k, iq k, jq 0q1 Step4: For a given siilarity criterion r, the ratio of siilar vectors of which distance is saller than r, denoted as A () r. i 1 ( ) ( ) Ai ( r) nu{ d[ Yk, ( i), Y k, ( j)] r},1 i P, i j (5) P1 where nu accuulates the nuber of siilar vectors. Step5: Define A () r as the average of A () r. i P 1 A ( r) Ai ( r) (6) P Step6: In this step, calculate A 1 ( f ). Different fro the accuulative function of saple entropy, we define a new accuulative function s here. s is a piecewise function that avoids the easureent of siilar vectors changing suddenly between 0 and 1. ( ) ( ) 0 d[ Yk, 1 ( i), Yk, 1( j)] f ( ) ( ) s d[ Yk, 1 ( i), Yk, 1 ( j)] (7) ( ) ( ) 1- d[ Yk, 1 ( i), Yk, 1( j)] f f where f is flexible siilarity criterion. The ratio of siilar vectors for pattern length +1 is as follow: P ( ) Ai f s (8) P1 Step7: Calculate A 1 ( f ) as follow: i1 j1 1 A ( f ) A ( f ) 1 P 1 i P i1 Step8: Calculate saple entropy for coarse-grained tie series 1 ( ) A ( f) k r f SapEn( y,,, ) ln[ ] A () r (10) Step9: Calculate flexible ultiscale entropy. 1 ( ) FMSE( x,,, r, f ) SapEn( y k,, r, f ) (11) k 1 Above is the calculation process of flexible ultiscale entropy. In the following calculations, we appoint =2, r=0.15,which eans siilarity criterion is 0.15*SD. SD is standard deviation of tie series{ x( i) 1 i N}. We appoint f=0.2, which eans flexible siilarity criterion is 0.2*SD Network Traffic Analysis Model Traditional ethods usually consider only one single space-tie scale, which are incapable of discovering soe inherent characteristics of network traffic. In this paper, we introduce a new ethod to analyze the spatioteporal characteristics of network traffic based on FMSE. First, we build the tie-space analysis odel as shown in Figure 1. Then, we select different space scales, tie scales and network levels, cut the original traffic data sequence into different sub sequences and y ( ) k (9) Copyright c 2016 SERSC 305

4 Network level International Journal of Grid and Distributed Coputing calculate the coplexity of each sequence with flexible utiscale entropy. Finally, suarize and copare the results of each sub sequence, discovering the evolveent echanis and the inherent characteristics of network traffic at different tie scales, space scales and network levels. We will conduct analysis fro two perspectives. One is acro level, and the other is pc level. We can first find a general law of network traffic fro acro level, and then zoo in to pc level to test the validity of the law. Since different individuals have different behaviors, soe particularities ay exist in pc level traffic. We will do a further analysis to discover the causes of particularities and find a general law that is suitable for all diensions of network traffic. Through analyzing network traffic fro two perspectives, we can iprove the accuracy and practicability of the law we discovered. Tie Figure 1. Tie and Space Analysis Model of Network Traffic 3. Experient Results 3.1. Experient Setup The data set used in our study was collected fro the east China s backbone of the CERNET (China Education and Research Network) and was released by Jiangsu Key Laboratory of Coputer Networking Technology on Noveber 9, The backbone covers over 100 universities and high schools. Its bandwidth was updated to 10G fro 2.5G in 2006 [12]. The duration of the data set is 1440 inutes and the size of the data set is GB. The analysis ethod introduced in section 2 is ipleented in Java language Macro Level Network Traffic Analysis At first, analyze network traffic fro acro level. We classify the original traffic data to different data sequences based on the network level, and then calculate the coplexity of each sequence by the flexible ultiscale entropy proposed in section 2. In the 306 Copyright c 2016 SERSC

5 calculation process, the pattern length is 2, the siilarity criterion is 0.15 and the flexible siilarity is 0.2. The axiu scale factor is often set to 20 in the calculation. However, considering that the length of traffic data contains points (seconds), we set the axiu scale factor to 180 and the scale interval to 5. Since the length of tie series is long enough, we can select different scale factors to analyze, so that it is easier to discover the inherent law of the tie series. Figure 2. Entropy Values Curve Based on Network Level Figure 2 is the flexible ultiscale entropy values of byte, packet, and flow series. In the figure, flow stands for traffic flow, which is based on the well-known five-tuple - the source IP address, destination IP address, source port, destination port and protocol fields. Pac stands for packet. The horizontal axis is the tie scale in seconds; the vertical axis is the value of flexible ultiscale entropy. The original data sequence is coarse-grained at scale factors fro 1 to 180 and calculated with flexible ultiscale entropy. Fro the figure, we can see the entropy value of traffic flow rises slowly as the scale factor increases, whereas the entropy value of packet and byte decreases slowly with the scale factor. Three entropy curves are relatively stable when the scale factor is ore than 45. Fro whole scale, the entropy value is in a relatively stable range, which indicates that network traffic exhibits long-range dependence characteristic. The larger the scale factor is, the ore inforation of the original data sequence is contained in the data sequence. Throughout all the scales, byte entropy is higher than packet and flow entropy, and the flow entropy is the lowest of the three. Furtherore, the relationship of coplexity trend of the three is relatively stable. The relative relationship aong byte, packet and traffic flow can be treated as an internal law of network traffic. And this law can be applied to analyze the behavior of network traffic, onitor and predict network traffic. The law will be further studied in the PC level traffic analysis PC level Network Traffic Analysis Based on the internal law of network traffic observed at acro level, we will investigate the terinal-level characteristics of network traffic in this section. In order to discover the law of coon characteristics in network traffic, we analyze the traffic generated by different sets of terinals. In the analysis, we use one hour s traffic data with a size of 78GB. We select eight network segents fro class A, B, C of the traffic data to analyze. The eight network segents contain 19, 26, 97, 2079, 44441, 50798, and terinals, respectively. These eight sets of data have a good continuity in an hour of tie series, and have an extensive coverage ranging fro 10 to 100 Copyright c 2016 SERSC 307

6 thousand terinals, and thus the results of the analysis are universal. By using analysis ethod in section 3.2, the traffic data of the eight groups were classified according to the network level, and calculate the flexible ultiscale entropy value for each data sequence. By coparing the coplexity of network traffic at different scales, finding the evolution echanis of network traffic at different teporal and spatial scales. 308 Copyright c 2016 SERSC

7 Figure 3. Entropy Value under Different Scales of Network Figure 3 shows the entropy values based on network level fro eight groups of traffic data at different scales. In the figure, flow stands for traffic flow, which is based on the well-known five-tuple - the source IP address, destination IP address, source port, destination port and protocol fields. Pac stands for packet. As can be seen fro Figure 3, for different size of eight groups of traffic data, the entropy value curve of flow is always lower than the packet entropy curve throughout all the scales. This shows that flow sequence coplexity is lower than the packet sequence coplexity, and the relative relationship between the packet and flow does not change as traffic scale changes. At the sae tie, last section s experient also shows that flow entropy is lower than packet entropy throughout all the scales. Whereas, in the eight figures, the coplexity relationship between byte and packet is not obvious in the whole scale, it changes in different scales. Then, we focus on the coplexity relationship between flow and byte sequence. Fro the figure, we can see the coplexity of the flow is higher than the coplexity of the bytes, except for 97 and terinal s entropy curves. And for the other six groups of data, the coplexity relationship between flow and byte is relatively stable. Over the whole scale, the coplexity of byte is higher than flow. In the rest of the paper, we focus on the particularity of the entropy curves of 97 and terinals. The above statistics is based on the nuber of bytes per second and the nuber of flow per second, however, the nuber of bytes include in each flow has a great difference. Fro the traffic data, we calculate the average bytes of a flow per second. By coparing the Copyright c 2016 SERSC 309

8 coplexity between flow sequence and the sequence of average bytes that one flow contained, discovering the coon coplexity relationship between flow and byte. We copare the coplexity of two kinds of sequence easured by flexible ultiscale entropy. As we copare the coplexity of two sequence directly, we only calculate the flexible ultiscale entropy at scale factor 1. Table 1. Entropy Value at Different Scale of Network Terinals Flow-Byte Entropy Flow Entropy Entropy Difference In Table 1, Flow-Byte Entropy stands for the flexible ultiscale entropy value of average bytes for one flow per second and Entropy Difference stands for the difference between the Flow-Byte Entropy and Flow Entropy. Fro Table 1, we can see that the entropy difference in the scale of 97 terinals and terinals is saller than 0. Due to the coplexity of flow-byte is lower than the coplexity of flow, the coplexity of flow is hence higher than the coplexity of byte. This phenoenon ay be due to any sall size flows or short duration flows in the network traffic. In order to verify the results, we investigate the nuber of average bytes including in one flow per second. Table 2. Relationship between Byte and Flow at Different Scale of Network Terinals Flow-Byte In Table 2, Flow-Byte stands for the nuber of bytes including in one flow per second. As show in the table 2, the bold line is the range of terinals and 97 terinals. We can see the values of flow-byte of these two scales are the sallest aong all the groups. The traffic of these two groups includes a larger nuber of sall flows than the others. The increase of sall flows leads to the increased coplexity of flow. The coplexity between flow and byte is changed as the relationship between flow and byte changes. Fro above results, we can see coplexity of flow is higher than the coplexity of byte in ost cases. 310 Copyright c 2016 SERSC

9 4. Conclusion and Future Work In this paper, we adopted flexible ultiscale entropy (FMSE). As an iproved ethod of ultiscale entropy, flexible ultiscale entropy has an iproved stability and accuracy of calculation, especially in large scale. We also proposed a ethod to analyze network traffic based on flexible ultiscale entropy, which is used as a tool to describe the coplexity of traffic sequence. By establishing the spatial and teporal characteristic network analysis odel, describing coplexity of network traffic clearly fro different diensions. Through the experient carried on one day s network traffic, of which the size is ore than 2TB. The result shows that there is a universal law aong the different levels of network traffic. For all scales, the coplexity of packet sequence is higher than that of flow sequence. In ost cases, the coplexity of byte sequence is higher than that of flow sequence. However, the coplexity of the flow sequence will be higher than that of byte sequence when there are any sall flows in network traffic. In our future work, we will go further to analyze network traffic and build a network situational awareness odel. We will use this odel to analyze network behavior, detect anoaly aong network and predict network traffic. Analyzing in real tie is also an iportant part of our future work. Acknowledgent The authors are grateful to the anonyous reviewers for their valuable coents and to the editors for their work that iproved this paper. This work was supported by NSF of Zhejiang under grant NO.LQ13F020017, LY14F020044, LY16F020018, and NSF of China under grant NO , , , and National Key Technology Research and Developent Progra of China under grant No.2014BAK14B04. References [1] W. E. Leland, M. S. Taqqu and W. Willinger, On the self-siilar nature of Ethernet traffic (extended version), IEEE/ACM Transactions on Networking, vol. 2, no. 1, (1994), pp [2] A. Y. Privalov and A. T. Analysis, Siulation of WAN Traffic by Self-Siilar Traffic Model with OMNET, Proceedings of 10th International Wireless Counications and Mobile Coputing Conference, Nicosia, Cyprus, (2014). [3] R. H. Riedi, M. S. Crouse and V. J. Ribeiro, A ultifractal wavelet odel with application to network traffic, IEEE Transactions on Inforation Theory, vol. 45, no. 3, (1998), pp [4] Y. Chen, B.Yang and Q. Meng, Sall-tie scale network traffic prediction based on flexible neural tree, Applied Soft Coputing, vol. 12, no. 1, (2012), pp [5] X. Liu, X. Fang and Z. Qin, A Short-Ter Forecasting Algorith for Network Traffic Based on Chaos Theory and SVM, Journal of Network & Systes Manageent, vol. 19, no. 19, (2011), pp [6] Q. F. Yao, C. F. Li, H. L. Ma and S. Zhang, Novel network traffic forecasting algorith based on grey odel and Markov chain, Journal of Zhejiang University, vol. 34, no. 4, (2007), pp [7] R. Xiang, J. Zhang and X. K. Xu, Multiscale characterization of recurrence-based phase space networks constructed fro tie series, Chaos, vol. 22, no. 1, (2012), pp [8] M. Costa, Multiscale entropy analysis of coplex physiologic tie series, Physical Review Letters, vol. 89, no. 6, (2002), pp [9] J. iihij rvi, M. Wellens and P. Mahonen, Measuring Coplexity and Predictability in Networks with Multiscale Entropy Analysis, Proceedings of IEEE INFOCOM 2009, Rio de Janeiro, Brazil, (2009). [10] J. S. Richan and J. R. Mooran, Physiological tie-series analysis using approxiate entropy and saple entropy, Aerican Journal of Physiology Heart & Circulatory Physiology, vol. 278, no. 6, (2000), pp. H2039-H2049. [11] S. D. Wu, C. W. Wu and S. G. Lin, Tie series analysis using coposite ultiscale entropy, Entropy, vol. 13, no. 3, (2013), pp [12] IPTRACE: Copyright c 2016 SERSC 311

10 Authors Chen Yang, he received the B.S degree in network engineering fro North University Of China, Taiyuan, China, in 2011 and M.S degree fro Hangzhou Dianzi University, Hangzhou, China, in His research interests include network security and coplex syste. Renjie Zhou, he is an assistant professor in School of Coputer Science and Technology at Hangzhou Dianzi University, Hangzhou, China. He received his Ph.D. degree fro Harbin Engineering University, Harbin, China, in He was a visiting scholar in the Departent of Electrical and Coputer Engineering at the University of Massachusetts at Aherst. Currently. His research interests include analysis of online social networks, and network security. JianWan, he is the Director of Grid and Service Coputing Lab in Hangzhou Dianzi University, Hangzhou, China, and is the Dean of the School of Coputer Science and Technology, Hangzhou Dianzi University. He received his Ph.D. degree fro Zhejiang University, Hangzhou, China, in His research areas include parallel and distributed coputing systes, virtualization, and grid coputing. He is a eber of the IEEE.. Jilin Zhang, he received the PhD degree in Coputer Applied Technology fro University of Science Technology Beijing, Beijing, China, in He is currently an associate professor at Hangzhou Dianzi University, China. His research interests include High Perforance Coputing and Cloud Coputing.. Yuyu Yin, he received the Ph.D. degree in coputer science fro Zhejiang University, Zhejiang, China, in He is currently an Associate Professor with Hangzhou Dianzi University, Hangzhou, China. His research interests include service coputing, cloud coputing, and iddleware techniques. 312 Copyright c 2016 SERSC