WNN-Based NGN Traffic Prediction

Transcription

1 WNN-Based NGN raffic Prediction Qigang Zhao, Xuming Fang, Qunzhan Li, Zhengyou He School of Electrical Engineering, Southwest Jiaotong University, Chengdu, Sichuan, 63,China Abstract In this paper we introduce a methodology to predict IP traffic in IP-based Next Generation Networ (NGN). By using Netflow traffic collecting technology, we ve collected some traffic data for the analysis from an NGN operator. o build wavelet basis Neural Networ (NN) we replace Sigmoid function with the wavelet in NN, and use wavelet multiresolution analysis method to decompose the traffic signal and then employ the decomposed component sequences to train the NN. By using the methods, we build a NGN traffic prediction model by which to predict one day s traffic. he experimental results show that the traffic prediction method of Wavelet NN(WNN) is more accurate than that without using wavelet in the NGN traffic forecasting.. Introduction Modeling and forecasting of NGN traffic is imperative to the support of multi-media applications with diverse statistical characteristics and Quality of Service (QoS) requirements. Due to the heterogeneity of NGN traffic such as the complex mixture of longrange and short-range dependence, it is difficult to use traditional models to analyze and predict the networ traffic since either they may not be able to capture complex dependence across times scales, or they may have very high computational complexity[-3]. Leland, aqqu, etc. have set the groundwor of considering self-similarity as an important notion in understanding the Internet traffic[-5]. Much of wor has been done to exploit the nonlinear correlation structure of IP networ. In [6], Xin Wang and Xiuming Shan use a wavelet-based method to predict Internet traffic. In [7], the author introduces a methodology to predict when and where lin additions/upgrades have to tae place in an IP bacbone networ. he article [8] and [9] respectively give a method to model heterogeneous networ traffic and a unified framewor for understanding IP traffic. Due to muti-services to be supported as well as different service having different traffic properties, the characteristics of NGN traffic are more difficult to capture and the traffic more difficult to predict than those of Internet though both of them are IP based. In view of the complex properties of NGN traffic, we give a novel method based on WNN to predict NGN traffic in this paper. In our method, the wavelet is used both to build transfer functions in NN forecast model and to decompose the traffic sequences into different frequency components in multiresolution analysis. In the analysis, the data are collected from the Integrated Service Networ (ISN) of Sichuan Unicom (NGN technology based). he paper is organized as follows. In the section, the method to collect the traffic data from the operator s routers based on Cisco Netflow protocol is presented. In section 3, the NGN IP traffic forecast model based on WNN is suggested and in section the experiment results are given. In the last section, the wor is concluded and the future tass are presented.. IP raffic Data Collection he IP traffic data adopted in this paper are collected from Sichuan Unicom s ISN (built with NGN technology). Sichuan Unicom s ISN provides voice, fax, data and video integrated service, with more than 5, customers at home subscribers. he maority of the networ s edge and core routers come from Cisco products. he edge routers are generally distributed and core routers equipped in /5/$. 5 IEEE. 3

2 communication center. he data are collected from both the edge (Cisco 73, 76) and the core routers (GSR). Figure. IP traffic collection model In Figure, we use a Unix worstation, laid in operator s Communication Center, as a server to collect traffic data from routers. he data transform between server and routers is based on NetFlow technology, which is advocated by Cisco and is a networ pacet switch technology. he techonolgy can be used to record networ flow information. One NetFlow, namely, a serial of data pacets from source to destination, is recorded, including information such as source and target IP addresses, transmit ports, types of protocols, types of services, and input interfaces. Netflow output parameters are defined at 8 routers, of which one is core router and the others are edge routers. he parameters to be set include output flow version, the number of flows, the size of output buffer, the IP addresses and port number of FlowCollector (server) etc. At FlowCollector, we configure receive port number, filter policy, and the directory to store flow files etc. he interval between flows is set at minutes and the data colleted covers a period of about months. Any equation: ψ t u us, () t = s ψ s. f L ( R) can be represented by the t u f () t = (, ) C ds + + Wf us ψ ψ s s du, s Where + Wf (,) u s f () t * t u dt s ψ = s. o discrete the integral variances s and u, the above equation can be replaced by the following expression: N N t b f() t = Wf( b, a) h N = = a. Generally let W = Wf ( b, a ), the signal can be approached by using the following wavelet, N t b St () = W h a = W, b, a St () herein, represent the coefficients of weight, translation and dilation respectively. he equation can be implemented with the neural networ (NN) as in Figure. Here, including in every neuron is not Sigmoid nonlinear function, but the wavelet function t b h. a 3. WNN based IP traffic forecast model 3.. Wavelet Basis NN Forecast Model A wavelet is a function average: + ψ () t dt = ψ L ( R) with a zero It is centered in the neighborhood of. A family of wavelet is obtained by scaling Ψ by s and translating it by u [] : Figure. Wavelet basis neural networ he best value of parameters can be W, b, a achieved by minimizing the energy function: E = [ S( t) S( t)] () 3

3 We choose the wavelet, and let = ( t b)/ a ht t t equation () can be expressed as follows: g( W) K = = [ S( t) S( t)] W cos.75 exp( / ) K ( ) = cos.75 exp( /), then the gradients of () g( b) K = = [ S( t) S( t)] WK [.75sin.75 b K (3).exp( )/ a ] + cos.75 exp( ) a g( a) K = = [ S( t) S( t)] WK [.75sin.75 a K exp( ). ] + cos.75 exp( ) a () hrough equation (), (3), (), we can calculate the W, b, a value of. 3.. Wavelet Based Signal Sequence Decomposing NGN IP traffic has the characteristics of being broad frequency domain, nonlinear and self-similar. he direct use of NGN traffic sequences in training the NN may mae the networ be unstable or tae much more time and data. For the purpose of getting stable networ and training the NN fast with fewer data, firstly we use wavelet to get different frequency sequences of IP traffic, and then mae use of different sequences to train the different subnets respectively. For signal St (), the sampling sequence is Sn ( ), n=,, Sn ( ),N. If represents the approximation of the signal at scale = ( c ( n) = S( n) ), the discrete wavelet transfer (DW) is described by the following equations. c ( n) = h( n) c ( ) + d ( n) = g( n) c ( ) + z z hn ( ) gn ( ) (5) In equations, the and are two serials of conugate filter coefficients determined by wavelet function ψ ( x). he scaling function is determined by the equation ψ() x = h ()( ϕ x ) (here, = x h () =< ϕ(), ϕ( x ) >and ψ() x = g ()( ϕ x ) ). = In equation (5), where the high-scale, low-frequency components c is the approximation and the low-scale, high-frequency components the detail of signal Sn ( ). hrough the decomposing of discrete signal c into d, d,, d, c from scale to J, the decomposed components contain the information of the signal from high-frequency to low-frequency WNN-Based NGN IP raffic Forecast Model As stated in the previous subsections, the steps to forecast the IP traffic of NGN are:. o gather several months of NGN IP traffic data;. o mae time-traffic sequences with the former data, and to decompose the sequences into different frequency components with DW. 3. o build the training sequences with every frequency component, and to train every sub- NN with the sequences;. o input the time variable t from the input layer of NN, we can get output forecast traffic in every output layer of sub-nn. o synthesize the output of all sub-nn, we get the traffic prediction at time t. he WNN-based traffic forecast model is demonstrated in Figure 3. Figure 3. WNN based traffic forcast model ( Notation: Sn ( ) raffic Sequences Data; SD: Signal Decomposing; H: High Frequency Filter; G: Low d 3

4 Frequency Filter;RC: Signal Reconstruct;M: Signal Synthesize.). Experimental Result Analysis he original time-traffic sequences curve is shown in figure -a. (For convenience, we choose about one day data.) Figure -b to figure -f demonstrate the decomposed sequence component A, D, D, D3 and D respectively, in which A is the approximation and the rest are the details Figure -a A Figure -c D Figure -e D Figure -b A Figure -d D Figure -f D Figure. Original & decomposed sequences Here, A = A + D+ D + D3+ D, and the low frequency component A contains most of the information of original signal A. o analyze the properties of all the components, we get the statistical characteristic value of the Component Sequence, shown as in able-. able. he statistical characteristic value Figure 5-a Original Signal Figure 5-b Prediction Result Figure 5-c Prediction Error Figure 5. WNN based signal prediction result o compare the prediction method, we directly employ the original signal sequences to train the non wavelet NN and to predict the last day s traffic. Figure 6-a tells the prediction result and Figure 6-b indicates the prediction error Figure 6-a Prediction Result Figure 6-b Prediction Error Figure 6. Non-wavelet NN based signal prediction result he errors comparison of the two methods are shown in table-. From the table, we can see the prediction accuracy is improved a lot by using WNN method. able. Prediction errors comparison able shows that the Standard Deviation (SD) of all components become smaller through wavelet decomposing. We use 9 days data to train the networ, and predict the last day s traffic. Figure 5-a suggests the original signal, 5-b the prediction result and 5-c the prediction error. (Notation: MRERR, Max Relative Error ; MXARER, Max Average Relative Error; RMSE: Average Square Root of Relative Error Square Sum; RC: Related Coefficient.) 33

5 5. Conclusions In this paper we venture to present a methodology for predicting IP traffic in integrated service networ NGN. In our model, we replace Sigmoid function with the wavelet to build wavelet basis NN, then use wavelets to decompose the traffic sequences into different frequency components and tae the components to train the sub-networs of Wavelet Basis NN. Our experimental analysis indicates that WNN can effectively improve the prediction accuracy compared with the NN without using wavelet. he factors that influence the traffic of NGN not only include the changing time, but also the holidays, momentous events, the service types that networ provides and the number of customers etc. he future wor remains for us is to exploit the relations between the above factors and the NGN IP traffic, and to build the traffic prediction model with the factors. 6. Acnowledgments We wish to than Operation Managing Department of Sichuan Unicom for its providing the IP traffic data of the NGN for the experiment. References [] Franlin D. Ohrtman, JR. SoftSwtich: Architecture for VoIP, he McGraw Companies INC [] Xusheng ian; Sheng Ma; Chuanyi Ji, Comparison of the independent wavelet models to networ traffic, Global elecommunications Conference,. GLOBECOM '. IEEE, 8 5,. [3] J. R. M. Hosing, Modeling persistence in hydrological time series using fractional differencing, Water Resources Res., vol., pp , 98. [] Norros, A storage model with self-similar input, Queuing Syst., vol. 6, pp , 99. [5] Xin Wang; Xiuming Shan, A wavelet-based method to predict Internet traffic,communications, Circuits and Systems and West Sino Expositions, 69-69,. [6] Papagiannai, K, "Long-term forecasting of Internet bacbone traffic: observations and initial models",infocom , 3 [7] Sheng Ma, Chuanyi Ji, Modeling heterogeneous networ traffic in wavelet domain, Networing, IEEE/ACM ransactions on, 63 69, 3. [8] Xusheng ian, He Wu, Chuanyi Ji, A unified framewor for understanding networ traffic using independent wavelet models, INFOCOM, 6 5,. [9] I. Daubechies, en Lectures on Wavelets. Philadelphia, PA: SIAM, 99. [] C.K. Chui, Wavelet: a tutorial in theory and application, New Yor: Academic,99. 3