Inerne Traffic Modeling for Efficien Nework Research Managemen Prof. Zhili Sun, UniS Zhiyong Liu, CATR UK-China Science Bridge Workshop 13-14 December 2011, London
Ouline Inroducion Background Classical eleraffic engineering model Main parameers of he Inerne races Well known mahemaical mehods New approach o he Inerne raffic daase Conclusion and direcions for furher sudies
Inroducion Hisorically, eleraffic engineering successful in elecommunicaion neworks (Poisson process Inerne raffic has grown significanly since 1990s; and over provision becomes impracical In 1990s, discovered ha he Poisson funcion failed o model he Inerne raffic. Many suggesed Pareo and self-similar models bu here is no conclusive confirmaion due o he complexiy of he Inerne raffic. This leaves a big gap beween he classical raffic engineering and he Inerne raffic modeling This paper presened a new approach
Teleraffic engineering componens Nework Conrol and managemen
Traffic load: E Basic conceps p p s 1 b b / p b Call Bandwidh (kbi/s 2 0.5 Erlang-Hour 3 2 Packe of 1 raffic uni 1 1 Erlang-Hour 1 Packe of 1 raffic uni Second 0.5 1 Hour 0 1 (a Traffic for elephony neworks (b Traffic for packe neworks
Arrival process Arrival ime of he i h packe is a T i as he following: 0 = T 0 < T 1 < T 2 <... < T i < T i+1 <.... For simpliciy, we can assume ha he observaion akes place a ime T 0 = 0. The number of calls in he inerval [0, is denoed as N. Here N is a random variable wih coninuous ime parameers and discree space. When increases, N never decreases. The ime disance beween wo successive arrivals is: X i = T i T i-1, i = 1, 2,... This is called he iner-arrival ime, and he disribuion of his process is called he inerarrival ime disribuion.
Number and Inerval represenaions Corresponding o he wo random variables N and X i, he wo processes can be characerized in wo ways: Number represenaion N : ime inerval is kep consan o observe he random variable N for he number of IP packes in. Inerval represenaion T i : number of arriving IP packes is kep consan o observe he random variable T i for he ime inerval unil here are n arrivals. The fundamenal relaionship beween he wo represenaions is given by he following simple relaion: N < n, if and only if, n = 1, 2,... This is expressed by Feller-Jensen s ideniy: Prob{N < n} = Prob{T n },
Exponenial and Poisson disribuions Three assumpions were made o model he arrival process using exponenial disribuion and Poisson disribuion: Saionary: For any arbirary 2 >0 and k0, he probabiliy ha k calls arrival in [ 1, 2 is independen of 1. Independence: The probabiliy of k calls arrival aking place in [ 1, 2 is independen of calls before 1. Simpliciy: i is call simple process if he probabiliy ha here is more han one calls arrival in a given poin of ime is 0.
The reasons for he failures of Poisson The main reasons are he naure of Inerne raffic and properies of TCP on which many applicaions are based including WWW, FTP, email, P2P, Telne, ec. These break he assumpions made for classic eleraffic engineering model, due o acknowledgemen, flow conrol and congesion conrol mechanisms. To invesigaed he feaures of he Inerne raffic o find alernaive raffic models and o show ha he Inerne raffic showed properies of long ail and self-similariy Bu sill can no fully model he real Inerne raffic, Due Inerne applicaions and heir complexiy,
Traffic parameers The raffic races conain informaion on each packe capured on he Inerne neworks. The flows of packes depend on he user aciviies and he applicaions used. The informaion in each packe capured includes: Time samp when he packe is capured Media Access Conrol (MAC frame header IPv4 or IPv6 Header Transmission conrol proocol (TCP header wih applicaion proocols such as HTTP, SMTP, FTP, ec. User daagram proocol (UDP header wih applicaion proocols such as DNS, RTP, ec.
Traffic races Traffic races observaions on 1 s Augus 2011, a a rans- Pacific line (150Mbps link in operaion since 1 s J uly2006: IPv4 packes couns 99.57% of he oal IP packes (99.6% in byes, Only 0.43% for IPv6 (0.4% in byes; i showed clearly ha he usage of IPv6 is sill very low,
TCP/UDP UDP for 16.81% (11.79% in byes For voice over IP, here is a consan sream of packes wih 14 byes of MAC header, 20 of IP, 8 of UDP and 12 of RTP; Plus payload of 160 byes for ITU-T G.711 codec as an example ha i has 64 kbps, 20 ms sample period and 1 frames per packe (20 ms [11]. TCP couns for 79.76% in packes (85.1% in byes; HTTP server couns for 35.04% in packes (64.78% in byes; HTTP clien couns for 20.36% in packes (7.2% in byes;
IP Traffic decomposiion 1 WWW Individual raffic models Measured IP raffic ( n i1 i n 2 3 4 5.. FTP Email VoIP. Ohers Modelled IP raffic
Exponenial Pareo Weibull Candidae Mahemaical funcions 0 0,, ( e f Disribuion funcion Densiy funcion 0,, 1 ( m m F 0 0,, 1 ( e F 0,, ( 1 m m f 0 0 ( 0,, 1 ( / ( for and F e F k 0 0 ( 0,, ( / ( 1 for andf e k k f k k
Fiing resuls Original whole raffic HTTP downlink Modeling he whole raffic Modeling for HTTP downlink raffic
Conclusion Due o he limiaion of classical echniques, difficul o o model he Inerne raffic. We inroduced a new approach o classify he mahemaical funcions using he reference funcion of g(= (1-z /, and Apply he funcion on he decomposed subse of he Inerne raffic raher han he complee daase Weibull disribuion gives a beer fiing han Pareo and exponenial funcions. Therefore, we can conclude ha he rang of disribuion funcions wih g(= (1-z /, where 0<= z <=1, provided he choices for modeling he decomposed Inerne raffic.
Direcions for furher sudies The resuls show ha here is a grea poenial for he new approach in he Inerne raffic engineering wih decomposiion of Inerne raffic. In fuure work, he new approach has ye been furher validaed for modeling on he decomposed componens of raffic, such as HTTP, FTP, Email, VoIP and Sreaming media, ec. The imporan issue remains: here is a new comprehensive model exis ha i is simple enough like he classical eleraffic engineering bu accurae enough for modeling he fuure Inerne raffic This paper presened a new approach o resolve he issues, hence am imporan opic for furher sudies.
Any Quesion? Z.Sun@surrey.ac.uk