Saint-Petersburg State University of Telecommunications Analysis of IP-oriented Multiservice Networks Characteristics with Consideration of Traffic’s Self-Similarity Properties Anatoly M. Galkin galkinam@inbox.ru Adviser: Dr., Professor Gennady G. Yanovsky Analysis of IP-oriented Multiservice Networks Characteristics with Consideration of Traffic’s Self-Similarity Properties OUTLINE •Why IP and why self-similarity? •Self-similarity, what is it? •Heavy-tailed Distributions •Self-similarity and Networks •Conclusions Analysis of IP-oriented Multiservice Networks Characteristics with Consideration of Traffic’s Self-Similarity Properties OUTLINE •Why IP and why self-similarity? •Self-similarity, what is it? NGN IP Traffic types •Heavy-tailed distributions •Self-similarity and Networks •Conclusions Why IP ? Growth of data services Channel switching Active introduction of IP networks Packet switching NGN – next generation network NGN is united network Supports different types of traffic Built on the base of the universal technology Divides switching, signaling and management Provides mentioned QoS (quality of service) Why IP ? voice NGN FR ... IP ATM -Data networks evolution to NGN: the problem of compatibility of technologies and standards (providing traffic transmission of different applications in united transport network) - Voice networks evolution to NGN: the problem of conversion from Channel Switching to Packet Switching Why IP ? 2001 year - Conceptual regulations about multiservice networks structure in Russian communication networks NGN architecture Management system Management Application servers Applications Softswitches Control Core Separate networks Access Packet network Mobile network UTRAN Business subscribers PSTN Broadband network DSL CS Remote office/SOHO Media Gateway WLL LE Home subscribers Mobile subscribers Why IP ? IP oriented networks Multiservice IP network applications classification of traffic types Type of traffic Applications IP telephony, videoconference Delay sensitivity Delay jitter sensitivity Low losses sensitivity RSVP, RTP, RTCP,UDP Control processes, on-line games Delay sensitivity Delay jitter sensitivity Losses sensitivity UDP, TCP Audio on demand Video on demand Internet broadcasting Low delay sensitivity Delay jitter sensitivity Losses sensitivity RSVP, SCTP, UDP,TCP Real time Stream Elastic Requirements Transport layer protocols Conference of documentation Low delay sensitivity Low delay jitter sensitivity High losses sensitivity Animation, file transfer, E-mail Very low delay sensitivity Low delay jitter sensitivity High losses sensitivity TCP Why self-similarity ? Problem of NGN is to provide QoS for all types of traffic QoS depends on service model Old Markovian models (memory-less), Poisson laws and Erlang formulas don’t work in new networks. 1993 year W. Lenard, M. Taqqu, W. Willinger, D. Wilson. “On the Self-Similar Nature of Ethernet Traffic” Analysis of IP-oriented Multiservice Networks Characteristics with Consideration of Traffic’s Self-Similarity Properties OUTLINE •Why IP and why self-similarity? •Self-similarity, what is it? Fractals distributions •Heavy-tailed Some mathematics •Self-similarity and Networks Hurst parameter •Conclusions Self-similarity, what is it? Fractals 1975 Benua Mandelbrot fractus (lat.)– consisting of fragments 0D 1D 2D 1.5D 3D Fern leaf Fractals property – self-similarity Fractals are determined by the equations of chaos Chaos deterministic chaos Stochastic fractal processes are described by selfsimilarity of statistical characteristics of the second order Self-similarity, what is it? Notations X ( X 1 , X 2 ,...) X ( m ) ( X 1( m ) , X 2( m ) ,...) Aggregated process Semi-infinite segment of second-order-stationary stochastic process Its discrete argument X t( m ) rm (k ) autocorrel ation function t N {1,2,...} Its parameters M [ X ] average D[ X ] 2 dispersion r (k ) r (k ) ( X t k )( X t ) 2 Let r(k) k-L1(k), k 1 X tmm1 ... X tm , m,t N m autocorrel ation function 0 1 L1 – is function slowly varying at infinity Self-similarity, what is it? Three definitions Process is 1.Exactly second-order self-similar (es-s) with the parameter H=1 ( / 2), 0< <1 If rm(k) = r(k), k Z+, m {2,3,…} 2.Second-order asymptotical self-similar (as-s) with the parameter H=1 ( / 2), 0< <1 r k g k , k N If mlim 3.Strictly self-similar (ss-s) with the parameter H=1 ( / 2), 0< <1 If m1-H X(m) = X, mN In other words X is es-s, if the aggregated process X(m) is indistinguishable from the initial process X at least in term of statistical characteristics in second order. X is as-s, if it meets es-s process after it is averaged on blocks of length m and m The relation between ss-s and es-s processes is analogous to relation between secondorder stationary process and strictly stationary process Self-similarity, what is it? Hurst parameter Harold Edwin Hurst detected that foodless and fertile years are not random 0<H<1 – Hurst parameter (exponent) H=0.5 – Brownian Motion 0<H<0.5 – antipersistence of the process 0.5<H<1 – persistent behaviour of the process or the process has long memory Analysis of IP-oriented Multiservice Networks Characteristics with Consideration of Traffic’s Self-Similarity Properties OUTLINE •Why IP and why self-similarity? •Self-similarity, what is it? •Heavy-tailed distributions Parameters and of distributions •Self-similarity Networks Heavy tails •Conclusions Pareto Weibull Log-normal Heavy-tailed distributions Probability distributions X – random value F(x)=P(X<x) – distribution function It determines probability of random value X<x, where x is certain value 0≤F(x)≤1 f(x)=dF(x)/dx – probability destiny f(x)≥0 M[x] – mathematical expectation M [ x] x f ( x)dx D[x] – dispersion, σ – root-mean-square deviation D[ x] ( x M [ x]) 2 f ( x)dx, ( x) D( x) C2 D[ x] M [ x]2 - quadratic coefficient of variation Heavy-tailed distributions Heavy-tailed distributions Self-similar processes could be described by so-called Heavy-tailed distributions Definition The random variable is considered to have heavy-tailed distribution if P[ Z x] ~ c x a , x with 0<a<2 a – shape parameter , c – a positive constant Light-tailed distributions (Exponential, Gaussian) have exponential decrease tails Heavy-tailed distributions have power law decrease tails 0<a<2 infinite dispersion 0<a≤1 also infinite average Network interest is the case 1<a<2 Then H=(3-a)/2 Heavy-tailed distributions Pareto distribution a b F ( x) P[ Z x] 1 , b x x a 1 1 0,9 0,7 ,b x a is the shape parameter, b is minimum value of x 0,8 0,6 0,7 0,5 0,6 0,4 0,5 0,4 0,3 0,3 0,2 0,2 0,1 0,1 0 0 0 2 4 6 Prob. density 8 10 12 14 Distribution function Pareto distribution is most frequently used (VoIP, FTP, HTTP) Distribution function a b b x 0,8 Prob. density f ( x) Pareto Distribution Heavy-tailed distributions Weibull distribution a Weibull Distribution , x x0 a , x x0 a is the shape parameter, β is the averaged weight speed x0 is the minimum value of x Prob. density F ( x) 1 e x x0 e x x0 0,6 1,2 0,5 1 0,4 0,8 0,3 0,6 0,2 0,4 0,1 0,2 0 0 0 1 2 Prob. density Weibull distribution is used for FTP 3 4 Distribution function 5 6 Distribution function a x x0 f x a 1 Heavy-tailed distributions Log-normal distribution 1 ln x m 2 , x 0 exp 2 f ( x) x 2 2 0, x 0 Log-normal Distribution 0,1 0,9 0,09 0,8 0,08 Prob. density 0,07 0,6 0,06 0,5 0,05 0,4 0,04 0,3 0,03 Distribution function 0,7 0,2 0,02 0,1 0,01 0 0 0 5 10 Prob. density 15 20 25 Distribution function It has a finite dispersion but has a subexponential decrease of a tail It used for call-centers, LANs, etc. Analysis of IP-oriented Multiservice Networks Characteristics with Consideration of Traffic’s Self-Similarity Properties OUTLINE Kendall classification Researches of networks •Why IP and why self-similarity? Limitations for real networks •Self-similarity, what is it? QoS parameters calculation •Heavy-tailed distributions Network modeling •Self-similarity and Networks •Conclusions Self-similarity and networks Kendall classification A/B/V/K/N Model of servicing 1 ... 1 N A(x) S K=S+V A – law of incoming traffic B – law of servicing traffic S – queue size V – number of severs K – number of places in system N – number of sources If N=∞ then A/B/V/K Often S=∞ → K=∞ then A/B/V 1 ... v B(x) Classic teletraffic models M/M/1, M/M/V/K , M/D/V etc. M – Poisson law F ( x) 1 e x D – determinate F(x)=const Self-similarity and networks 1993 year W. Lenard, M. Taqqu, W. Willinger, D. Wilson. “On the Self-Similar Nature of Ethernet Traffic” The period is 4 years From 3 pieces of Bellcore network It has been shown that 0.7<H<0.98 Poisson Measured Further researches Now – about 10000 works about self-similarity M. Taqqu, W. Willinger, K. Park, M. Crowell - research on the network layer. W. Willinger, M. Taqqu, R. Sherman, D. Wilson, A. Erramili, O. Narayan - research of the Ethernet traffic on data link layer N. Sadek, A. Khotanzad, T. Chen - the АТМ traffic K. Park, G. Kim, M. Crovella,V. Almeida, A. de Oliveira, A. B. Downey - research of TCP applications In S. Molnar’s paper VoIP traffic is observed Researches in Russia The interest to self-similarity in Russia was initiated by V.I. Neiman Rigorous mathematics description of self-similar processes is given by B. Tsibakov Applications of self-similar processes in telecommunications are presented in the book written by O. Sheluhin Another works by A.J. Zaborovski, V.S. Gorodetski, V.V. Petrov Self-similarity and networks Further researches DISTRIBUTION LAWS FOR DIFFERENT TYPES OF TRAFFIC IN IP NETWORKS A is law of incoming traffic B is law of of size of protocol data blocks Traffic type VoIP FTP/TCP SMTP/TCP Distribution law M is Poisson law Authors А В P Р Molnar P W and LN Van Mieghem Downey М P is Pareto law М Molnar HTTP/TCP P LN and P Crovella Van Mieghem IP P P Paxson Ethernet P P Taqqu ATM D F-ARIMA Sadek LN is lognormal law F-ARIMA is Fractal Auto-regressive Integrated moving Average D is determinate Self-similarity and networks Even if one source generate self-similar traffic then aggregated traffic has self-similar properties. At the network layer aggregated traffic is described with P/P/m most adequately Self-similarity and networks Insertion of limitation for real values of random quantities If random value is the size of protocol data block then turndown of value is [k; L]. k is minimum size L is maximum. Restricted distribution Pareto Distribution 0,8 0,7 Prob. density 0,6 0,5 0,4 L 0,3 0,2 0,1 0 0 2 4 6 8 Prob. density 10 12 14 Self-similarity and networks Insertion of limitation for real values of random quantities Restricted distribution has a finite parameters Mx and Dx 2 - finite value Then C 2 2 Mx For Pareto law Lk L k M x 1 L k L k 2 2 L2 k L k 2 Lk L k 2 2 1 L k 2 1 L k L2 k L k 2 Lk L k 2 C 2 1 2 L k Lk L k 2 2 Self-similarity and networks Now we could calculate QoS parameters – delays and losses Delays Losses Ca2 Cs2 P , m 1 2 2 1 Ploss t m t t s 2 2 2 1 Ca Cs ts C a2 C s2 t P , m m1 2 2 2 Ca Cs nb nb1 nb – buffer size - system load - average time of the packet’s service t - average time of the packet’s staying ts Ca2 and C s2 are quadratic coefficients of in the buffer. variation of incoming flow and service time - average value of the packets’ number in the queue distributions, correspondingly tm - average value of delay parameter m P,m m1 1 m Self-similarity and networks Graphics Loss probability in P/G/1 system for different distributions of service time The average delay in P/G/1 system for different distribution laws of service time Self-similarity boils down to packet losses, delays and congestions Self-similarity and networks Multiservice traffic modeling Excel, MathCAD, MathLAB – non specialized OPNET, COMNET ect. GPSS General Purpose Simulating System Allows to research discrete models of different types NS2 network simulator 2 Object-oriented discrete event simulator. Useful for simulating local and wide area networks The main advantage – it is free !!! ns2 Network simulator 2 (ns2) 1996 year Project VINT (Virtual InterNetwork Testbed), organized by DARPA (Defense research project agency) •Specialized for existing modern technologies •Open source code software •Core modification availability •Ns2 is free product •Result visualization availability Pareto TCP 2Mb TCP1sink TCP2sink 2Mb Pareto TCP FIFO RED 2Mb 2Mb FIFO RED 2Mb Pareto UDP UDPnull Results of modeling • • Animation Trace file Ploss for P/P/m Ploss analysis results 0,3 modeling 0,25 0,2 0,15 0,1 0,05 0 0 0,2 0,4 0,6 0,8 Conclusions •NGN is based on multiservice IP-oriented network •Providing QoS is one of the main problem •Multiservice IP traffic has a self-similarity properties •Old distributions (Poisson) don’t work •IP-traffic has Heavy-tailed distributions (the main is Pareto) •Self-similarity makes worse QoS parameters THANK YOU !