Beruflich Dokumente
Kultur Dokumente
Vaton
Trafc Models
Sandrine VATON {sandrine.vaton@telecom-bretagne.eu}
ecom Tel Bretagne
Goals :
to analyze trafc measurements (le sizes, trafc volumes/min., packet sizes (in bytes), packet interarrival times, packet headers, etc...), to identify some characteristic properties (correlation, laws of distribution, etc...) to propose some trafc models, most of the time statistical models, that take into account the main properties discovered in the measured trafc, knowledge : statistics, statistical processes, signal processing, time series ...
Desired skills :
to clarify some denitions (long-range dependence, self-similarity, etc...) to detect some properties in the trafc (heavy tails, long-range dependence, etc...) parameter estimation : the value of the parameters of a model are estimated from trafc measurements to produce synthetic trafc from theoretical trafc models to understand the impact of some characteristic properties (e.g. : large buffer asymptotics for a queue when the input trafc is self-similar)
Trafc Models S. Vaton Introduction Exponential Distribution, Poisson Process Continuous Time Markov Chains Markov Queues Phase-type distributions MAP Processes BMAP Processes
Trafc Models S. Vaton Introduction Exponential Distribution, Poisson Process Continuous Time Markov Chains Markov Queues Phase-type distributions MAP Processes BMAP Processes
Markovian Models
1 2 3 4 5 6 7
Introduction Exponential Distribution, Poisson Process Continuous Time Markov Chains Markov Queues Phase-type distributions MAP Processes BMAP Processes
Trafc Models S. Vaton Introduction Exponential Distribution, Poisson Process Continuous Time Markov Chains Markov Queues Phase-type distributions MAP Processes BMAP Processes
markovian models : from the Poisson model (Erlang, 1907) to BMAP processes (years 1980s) markovian models
very usual trafc models relatively simple usage (closed-form performance formulas) justify Markov models of queues solved problems : parameter estimation, performance of queues when the input trafc is Markov
Trafc Models S. Vaton Introduction Exponential Distribution, Poisson Process Continuous Time Markov Chains Markov Queues Phase-type distributions MAP Processes BMAP Processes
Trafc Models S. Vaton Introduction Exponential Distribution, Poisson Process Continuous Time Markov Chains Markov Queues Phase-type distributions MAP Processes BMAP Processes
Consequences :
standard markovian models are inaccurate QoS predictions (delay, loss, etc...) that are based on markovian models are dramatically optimistic it becomes necessary to :
design new trafc models (heavy tails, self-similarity, long-range dependence, etc...) Pareto distribution, Weibull distribution, -stable distribution, fractional Brownian motion (fBm), fractional Gaussian noise (fGn), fractional ARIMA process (fARIMA), etc... parameter estimation for these new models performance evaluation for these new models (queues), trafc generators, etc...
Trafc Models S. Vaton Introduction Exponential Distribution, Poisson Process Continuous Time Markov Chains Markov Queues Phase-type distributions MAP Processes BMAP Processes
Phase-type distributions :
they generalize the exponential distribution, coefcient of variation cv > 1 or cv < 1, but several parameters more general, but more complicated association of several exponential distributions in series or in parallel (several phases) Erlang, Hypo-Exp, Hyper-Exp, Cox, etc...
Trafc Models S. Vaton Introduction Exponential Distribution, Poisson Process Continuous Time Markov Chains Markov Queues Phase-type distributions MAP Processes BMAP Processes
Trafc Models S. Vaton Introduction Exponential Distribution, Poisson Process Continuous Time Markov Chains Markov Queues Phase-type distributions MAP Processes BMAP Processes
MAP
BMAP
MMPP
PH
MMPP2
COX
HypoEXP HyperEXP
IPP
Erlang
EXP
Trafc Models S. Vaton Introduction Exponential Distribution, Poisson Process Continuous Time Markov Chains Markov Queues Phase-type distributions MAP Processes BMAP Processes
P(x1 X x2 ) =
p(u) du
x1
p(u) du
p(u) du
x
moments E (X n ), n = 1, 2, 3, . . . , moment generating function M ( ) = E(exp(X )), cumulant generating function ( ) = log M ( )
Trafc Models S. Vaton Introduction Exponential Distribution, Poisson Process Continuous Time Markov Chains Markov Queues Phase-type distributions MAP Processes BMAP Processes
Trafc Models S. Vaton Introduction Exponential Distribution, Poisson Process Continuous Time Markov Chains Markov Queues Phase-type distributions MAP Processes BMAP Processes
Exponential distribution : 1st and 2nd order cumulants let us consider a random variable X Exp() :
mean [1st order cumulant] : E(X ) = 1/, (unit of : sec1 ), variance [2nd order cumulant] : var(X ) = E(X 2 ) (E(X ))2 = (1/)2 , standard deviation : std(X ) =
var(X ) = 1/,
Trafc Models S. Vaton Introduction Exponential Distribution, Poisson Process Continuous Time Markov Chains Markov Queues Phase-type distributions MAP Processes BMAP Processes
Memoryless Property
Memoryless property : let X be a continuous R.V. with exponential distribution, X Exp() ; then the memoryless property holds : P(X x + u | X x) = P(X u), x 0, u 0 it means that if X represents a duration for example, the time spent since start time does not change anything on the distribution of the time left before the end ; typically in the framework of queuing theory, X could be the service time or the interarrival time (time between arrivals of two consecutive clients) this property is fundamental since it justies the solution of some queues as Markov chains (e.g. queues with Poisson input and exponential service times)
Trafc Models S. Vaton Introduction Exponential Distribution, Poisson Process Continuous Time Markov Chains Markov Queues Phase-type distributions
E3
E4
E5
T2 T3 arriveearriv ee
client 2 client 3
T4 arrivee
client 4
T5 arrivee
client 5
T6 Temps arrivee
client 6
5 4 3 2 1 0
Temps t
The count process N (t) counts, for example, the number of arrivals of clients on [0, t].
Trafc Models S. Vaton Introduction Exponential Distribution, Poisson Process Continuous Time Markov Chains Markov Queues Phase-type distributions MAP Processes BMAP Processes
stationary increments : the distribution of N (t + ) N (t) (0 < t, ) is the same as the distribution of N ( ) N (0), N (t + ) N (t) is a Poisson R.V. with mean P(N (t + ) N (t) = k) = exp( ) ( )k k!
is the average number of events (arrival of clients, etc...) per time unit (unit of : sec1 ).
Trafc Models S. Vaton Introduction Exponential Distribution, Poisson Process Continuous Time Markov Chains Markov Queues Phase-type distributions MAP Processes BMAP Processes
Poisson process : exponential interarrival times Let {N (t), t R} be a count process, let T1 T2 T3 . . . be the arrival times, and let E1 = T2 T1 , E2 = T3 T2 , E3 = T4 T3 , . . . be the corresponding interarrival times. The following proposals are equivalent : E1 , E2 , E3 , . . . are independent and identically distributed R.V. with distribution Exp(). {N (t), t R} is a Poisson process with rate ,
Trafc Models S. Vaton Introduction Exponential Distribution, Poisson Process Continuous Time Markov Chains Markov Queues Phase-type distributions MAP Processes BMAP Processes
6 5 4 3 2 1 0
t0
t1 t2 t3
t n 1 t n
Trafc Models S. Vaton Introduction Exponential Distribution, Poisson Process Continuous Time Markov Chains Markov Queues Phase-type distributions MAP Processes BMAP Processes
Interpretation of the weak Markov property future values of the process Xt depend on the present value only the only memory is the current value of the process past and future are independent conditionnally to the present value of the process
Trafc Models S. Vaton Introduction Exponential Distribution, Poisson Process Continuous Time Markov Chains Markov Queues Phase-type distributions MAP Processes BMAP Processes
Characterization of a continuous time discrete state Markov chain A continuous time discrete state Markov chain can be characterized by : its state transition diagram or, equivalently by its innitesimal generator
Trafc Models S. Vaton Introduction Exponential Distribution, Poisson Process Continuous Time Markov Chains Markov Queues Phase-type distributions MAP Processes BMAP Processes
0.1 1.2 0 0
0.3 0 0 1.9
Trafc Models S. Vaton Introduction Exponential Distribution, Poisson Process Continuous Time Markov Chains Markov Queues Phase-type distributions MAP Processes BMAP Processes
For example, if the process is in state 1 then it remains in state 1 for a time distributed as an exponential R.V. with parameter 0.5 = 0.1 + 0.3 + 0.1 (average value 1/0.5 = 2 seconds) and after that time the process changes its state for state 2 with proba. 0.1/0.5 = 1/5 or state 3 with proba. 0.3/0.5 = 3/5 or state 4 with proba. 0.1/0.5 = 1/5.
Trafc Models S. Vaton Introduction Exponential Distribution, Poisson Process Continuous Time Markov Chains Markov Queues Phase-type distributions MAP Processes BMAP Processes
Innitesimal Generator
The innitesimal generator of this example is the following matrix : 0.5 0.1 0.3 0.1 0.8 1.2 0 0.4 Q= 0 0 0 0 0 0 1.9 1.9
non-diagonal elements are transition rates : qij is the rate of transition from state i to state j (i = j ) non-diagonal elements are always positive or null diagonal elements are always negative or null the sum of the elements over each row is always 0 ; qii = j =i qij
Trafc Models S. Vaton Introduction Exponential Distribution, Poisson Process Continuous Time Markov Chains Markov Queues Phase-type distributions MAP Processes BMAP Processes
Markov Queues
Markov Queues if the following conditions are satised :
1
the arrival of clients is a Poisson process (i.e. the interarrival times are independent and identically distributed as Exp()) the service times are distributed as Exp()
then the number of clients in the system is a continuous time discrete states Markov chain the weak Markov property holds for the number of clients in the system because the memoryless property is satised by the interarrival times and the service times the theory of Markov chains is a powerful framework for solving some performance evaluation problems (typically queues with Poisson arrivals and exponential service times)
Trafc Models S. Vaton Introduction Exponential Distribution, Poisson Process Continuous Time Markov Chains Markov Queues Phase-type distributions MAP Processes BMAP Processes
Example : M/M/1
Example of the M/M/1 :
M/M/1 arrival of clients : Poisson process with rate (average number of clients per second) service times : Exponential distribution with parameter (average service time : 1/ seconds) 1 server no limitation on the size of the queue (innite buffer) FIFO : First In First Out
Trafc Models S. Vaton Introduction Exponential Distribution, Poisson Process Continuous Time Markov Chains Markov Queues Phase-type distributions MAP Processes BMAP Processes
Trafc Models S. Vaton Introduction Exponential Distribution, Poisson Process Continuous Time Markov Chains Markov Queues Phase-type distributions MAP Processes BMAP Processes
Trafc Models S. Vaton Introduction Exponential Distribution, Poisson Process Continuous Time Markov Chains Markov Queues Phase-type distributions MAP Processes BMAP Processes
Performance Measures : performance measures (input/output rate, server utilization, average delay, etc...) are easily obtained from the steady state distribution (n) input/output rate : server utilisation : average delay :
1 1 1
1 1
Trafc Models S. Vaton Introduction Exponential Distribution, Poisson Process Continuous Time Markov Chains Markov Queues Phase-type distributions MAP Processes BMAP Processes
M/M/C/C
M/M/C/C
lost
2
M/M/C/C :
arrival of clients : Poisson process with rate service times : Exponential distribution with rate C servers, no buffer if a client arrives and if a server is free then the client starts service immediately ; otherwise the client is lost
Trafc Models S. Vaton Introduction Exponential Distribution, Poisson Process Continuous Time Markov Chains Markov Queues Phase-type distributions MAP Processes BMAP Processes
State Transition Diagram : the number of clients in the system (service or buffer) is a CTMC with the following state transition diagram
0 1
C1
4 (C 1) C
Trafc Models S. Vaton Introduction Exponential Distribution, Poisson Process Continuous Time Markov Chains Markov Queues Phase-type distributions MAP Processes BMAP Processes
0nC
Trafc Models S. Vaton Introduction Exponential Distribution, Poisson Process Continuous Time Markov Chains Markov Queues Phase-type distributions MAP Processes BMAP Processes
Phase-type distributions
Phase-type distributions PH-type distributions generalize the exponential distribution PH-type distribution = association of n exponential distribution in series or in parallel (n phases) more general than the Exponential distribution (in particular, they make it possible to take into account some cases when Cv = 1) but more parameters than the Exponential distribution Neuts, Matrix Geometric Solutions in Stochastic Models, Johns Hopkins University Press, 1981 popular cases of PH-type distributions : Erlang, Hypo-Exponential, Hyper-Exponential, Cox, etc...
Trafc Models S. Vaton Introduction Exponential Distribution, Poisson Process Continuous Time Markov Chains Markov Queues Phase-type distributions MAP Processes BMAP Processes
Erlang-distribution
F IG .: Erlang-k Distribution
X = X1 + X2 + . . . + Xk with Xi Exp()
Trafc Models S. Vaton Introduction Exponential Distribution, Poisson Process Continuous Time Markov Chains Markov Queues Phase-type distributions MAP Processes BMAP Processes
X = X1 + X2 + . . . + Xk with Xi Exp(i )
Ho(k) distribution = k Exponential Phases in series Erlang-k distribution is a particular case of Ho(k) distribution (when 1 = 2 = . . . = k ) in the case of the Ho(k) distribution it holds that Cv < 1, hence the name hypo-exponential
Trafc Models S. Vaton Introduction Exponential Distribution, Poisson Process Continuous Time Markov Chains Markov Queues Phase-type distributions MAP Processes BMAP Processes
pk
Hr(k) distribution = k Exponential phases in parallel Cv > 1 hence the name Hyper -exponential
p2
Trafc Models S. Vaton Introduction Exponential Distribution, Poisson Process Continuous Time Markov Chains Markov Queues Phase-type distributions MAP Processes BMAP Processes
Coxian distribution
1
1
2 a1
2
3 a2
3 ... ... ... ... k1
k ak 1
k
1 a1
1 a2
1 ak 1
Trafc Models S. Vaton Introduction Exponential Distribution, Poisson Process Continuous Time Markov Chains Markov Queues Phase-type distributions MAP Processes BMAP Processes
Trafc Models S. Vaton Introduction Exponential Distribution, Poisson Process Continuous Time Markov Chains Markov Queues Phase-type distributions MAP Processes BMAP Processes
Examples (1/2)
1
1/3
2/3
Characterization by the following CTMC : state space {e, 1, 2, 3} (e, absorbing state) initial distribution : = (1/3, 0, 2/3) for the transient states {1, 2, 3} and 0 for the absorbing state e innitesimal generator 1 0 0 0 0 0 C B 0 1 1 0 C Q=B A @ 2 0 2 0 3 0 0 3
Trafc Models S. Vaton Introduction Exponential Distribution, Poisson Process Continuous Time Markov Chains Markov Queues Phase-type distributions MAP Processes BMAP Processes
Examples (2/2)
Exercice : In the case of Ho(k), Hr(k), and Co(k) distributions please nd out the CTMC that characterizes the distribution : state space, state transition diagram, initial distribution, innitesimal generator.
Trafc Models S. Vaton Introduction Exponential Distribution, Poisson Process Continuous Time Markov Chains Markov Queues Phase-type distributions MAP Processes BMAP Processes
Characterization
innitesimal generator of the corresponding CTMC 0 0 0 ... ... 0 0 t1 t2 t3 T Q= . . . . . . tk where T is a matrix of size k k which characterizes the transition rates between transient states t is a vector of size k 1 such that t = T e where e = [1 . . . 1] initial distribution : , vector of size 1 k which characterizes the initial distribution (the absorbing state is omitted)
Trafc Models S. Vaton Introduction Exponential Distribution, Poisson Process Continuous Time Markov Chains Markov Queues Phase-type distributions MAP Processes BMAP Processes
Temps
F IG .: Correlated Arrivals
MAP processes = generalization of PH-type distributions contrary to PH-type distributions, MAP processes take into account the correlations between the successive interarrival times a popular particular case of MAP = the Markov Modulated Poisson Process (MMPP)
Trafc Models S. Vaton Introduction Exponential Distribution, Poisson Process Continuous Time Markov Chains Markov Queues Phase-type distributions MAP Processes BMAP Processes
Markov Modulated Poisson Process (MMPP) MMPP (Markov Modulated Poisson Process)
MMPP = Poisson process in which the rate depends on the state of a CTMC with nite state space
CMTC
MMPP(2)
Trafc Models S. Vaton Introduction Exponential Distribution, Poisson Process Continuous Time Markov Chains Markov Queues Phase-type distributions MAP Processes BMAP Processes
Markov Modulated Poisson Process MMPP = particular case of Doubly Stochastic Poisson Process (DSPP)
DSPP = Poisson process in which the rate (t) is itself a stochastic process Poisson DSPP
) P(N (t) = k ) = exp(t) (t k!
k
t 0
(u) du)
R ( 0t (u) du)k k!
MMPP = Poisson process for which the rate (t) is a CTMC with states {1 , 2 , . . . , k }
Trafc Models S. Vaton Introduction Exponential Distribution, Poisson Process Continuous Time Markov Chains Markov Queues Phase-type distributions MAP Processes BMAP Processes
Markovian Arrival Processes (MAP) MAP = generalization of PH-type distributions correlation between successive interarrivals (contrary to PH-type distributions)
once the absorption has occurred the Markov chain starts again from one of the transient states, taking into account the last visited transient state the distribution of the state from which the Markov chain starts again depends on the last visited transient state
this distribution was noted in the case of PH-type distributions ; in the case of MAP processes there is one distribution (i), i = 1, 2, . . . , k for each transient state i {1, 2, . . . , k})
Trafc Models S. Vaton Introduction Exponential Distribution, Poisson Process Continuous Time Markov Chains Markov Queues Phase-type distributions MAP Processes BMAP Processes
P oisson(1 ) P oisson(2 )
` 2 etats F IG .: MMPP a
Trafc Models S. Vaton Introduction Exponential Distribution, Poisson Process Continuous Time Markov Chains Markov Queues Phase-type distributions MAP Processes BMAP Processes
1
e
1 0 0 A (1 + ) ( + ) 2 1 t= 2 (1 + ) T = ( 2 + ) 0 0 @ 1 2
2 initial distributions : 1 = (1, 0) if the absorption has occurred from state 1 ; 2 = (0, 1) if the absorption has occurred from state 2
Trafc Models S. Vaton Introduction Exponential Distribution, Poisson Process Continuous Time Markov Chains Markov Queues Phase-type distributions MAP Processes BMAP Processes
BMAP
Trafc Models S. Vaton Introduction Exponential Distribution, Poisson Process Continuous Time Markov Chains Markov Queues Phase-type distributions MAP Processes BMAP Processes
Example : batch MMPP(2), (1=red ; 2=green ; 3=blue) CTMC with 2 states when the CTMC is in state 1, the packets are produced as a Poisson(1 ) process, packets are red with proba. = 1/3, green with proba.=1/3, blue with proba.=1/3 when the CTMC is in state 2, the packets are produced as Poisson(2 ), packets are red with proba. = 1/2, green with proba. = 0, blue with proba. = 1/2
Trafc Models S. Vaton Introduction Exponential Distribution, Poisson Process Continuous Time Markov Chains Markov Queues Phase-type distributions MAP Processes BMAP Processes
MAP
BMAP
MMPP
PH
MMPP2
COX
HypoEXP HyperEXP
IPP
Erlang
EXP
Trafc Models S. Vaton Heavy-Tailed Distributions Long-Range Dependence Self Similarity Agregation Models Impact on Performance
Trafc Models S. Vaton Heavy-Tailed Distributions Long-Range Dependence Self Similarity Agregation Models Impact on Performance
Paxson and Floyd, The failure of Poisson modelling, 1995 interrarival times are not exponential, bursts over several time scales, slowly-decaying correlations (long-range dependence), heavy tails (in duration/volume of TCP connections ; le sizes, transfer times, reading times in WWW, etc...), self-similarity of the cumulated workload, etc...
Trafc Models S. Vaton Heavy-Tailed Distributions Long-Range Dependence Self Similarity Agregation Models Impact on Performance
Goals of this class clarify the concepts of heavy tails, long-range dependence, self-similarity introduce some tests to detect long-range dependence, etc... in trafc introduce some trafc models with heavy tails, long-range dependence, self similarity, etc... temptative explanation of the source of long-range dependence in trafc study of the impact of long-range dependence on QoS (delays, buffer occupation, etc...)
Trafc Models S. Vaton Heavy-Tailed Distributions Long-Range Dependence Self Similarity Agregation Models Impact on Performance
Heavy-Tailed Distributions Long-Range Dependence Self Similarity Agregation Models Impact on Performance
10
11
12
Trafc Models S. Vaton Heavy-Tailed Distributions Long-Range Dependence Self Similarity Agregation Models Impact on Performance
where F c (t) = P(X t) is the complementary cumulative distribution function. Examples of heavy-tailed distributions :
t Pareto distribution : F c (t) = ( tmin ) with tmin , > 0 (power-law tailed) ; innite variance if < 2 ; innite mean, innite variance if 1 Weibull distribution : F c (t) exp((t/a)c ) with c < 1 and a > 0 ; heavy-tailed but not power-law tailed
Trafc Models S. Vaton Heavy-Tailed Distributions Long-Range Dependence Self Similarity Agregation Models Impact on Performance
power-law tails : the complementary cumulative distribution function is linear in log-log scale
Trafc Models S. Vaton Heavy-Tailed Distributions Long-Range Dependence Self Similarity Agregation Models Impact on Performance
12
10
Traffic Volume
20
40
60
80
100
120
Time (unit=hour)
Trafc Models S. Vaton Heavy-Tailed Distributions Long-Range Dependence Self Similarity Agregation Models Impact on Performance
-stable Distributions
-stable distributions : often used in order to model very impulsive phenomenons (bursts) in what follows we restrict ourselves to -stable symetric distributions the distribution is approximately gaussian around the mean value but, when < 2, the -stable distribution is heavy-tailed (more precisely, power-law tailed)
Trafc Models S. Vaton Heavy-Tailed Distributions Long-Range Dependence Self Similarity Agregation Models Impact on Performance
4
20
3
15
2
10
1
0
2
5
3
10
50
100
150
200
250
300
350
400
450
500
15
0
4
50
100
150
200
250
300
350
400
450
500
150
12
x 10
10
100
50
50
100
50
100
150
200
250
300
350
400
450
500
50
100
150
200
250
300
350
400
450
500
The phenomenon is more and more impulsive for small values of the parameter .
Trafc Models S. Vaton Heavy-Tailed Distributions Long-Range Dependence Self Similarity Agregation Models Impact on Performance
Trafc Models S. Vaton Heavy-Tailed Distributions Long-Range Dependence Self Similarity Agregation Models Impact on Performance
2 important particular cases of -stable distributions Gaussian distribution if = 2 : f=2,, (x) = (x )2 1 ) exp( 4 4
Trafc Models S. Vaton Heavy-Tailed Distributions Long-Range Dependence Self Similarity Agregation Models Impact on Performance
-stable distributions with < 2 behave as a power-law in the tails of the distribution : limt t F c (t) = cst. inuence of the Levy index :
the smallest the value of (0 < < 2) the strongest is the phenomenon of heavy tails for small values of extremely large values of the R.V. are not the exception ; modelisation of very impulsive phenomenons
Trafc Models S. Vaton Heavy-Tailed Distributions Long-Range Dependence Self Similarity Agregation Models Impact on Performance
Moments of -stable distributions if < 2 then the variance of X is innite if moreover 1 then the mean of X is also innite
Trafc Models S. Vaton Heavy-Tailed Distributions Long-Range Dependence Self Similarity Agregation Models Impact on Performance
-stable distributions : generation methods we restrict ourselves to the case of symetric -stable distributions even if similar expressions exist in the general case generation of a symetric -stable distribution W, exponential R.V. with parameter = 1 ; , R.V. with uniform distribution over ] /2, +/2[ X= sin() cos((1 )) 1 ) ( W (cos())1/
Trafc Models S. Vaton Heavy-Tailed Distributions Long-Range Dependence Self Similarity Agregation Models Impact on Performance
2 important particular cases : = 2, Gaussian distribution X= sin(2) cos() 1/2 ) = 2 W sin() ( 1 / 2 W (cos())
Box-Muller algorithm for the generation of the N(0, 1) distribution. = 1, Cauchy distribution X = tan() This is a classical method for the generation of the Cauchy distribution.
Trafc Models S. Vaton Heavy-Tailed Distributions Long-Range Dependence Self Similarity Agregation Models Impact on Performance
Paxson et Floyd, The failure of Poisson modelling long-term correlations in trafc this long-term correlation is contradictory with classical (Markovian) trafc models (Poisson, MMPP, MAP, etc...) Long-Range Dependence (LRD) denition signature (autocovariance function, power density spectrum, variance-time analysis) some examples of LRD processes (fractional Gaussian noise, fARIMA) generation of LRD processes
Trafc Models S. Vaton Heavy-Tailed Distributions Long-Range Dependence Self Similarity Agregation Models Impact on Performance
Spectral Representation
2nd order stationarity (stationarity of the covariances) Let Xt be a discrete time stochastic process (time series). This process is 2nd order stationary if : E(Xt ) = m does not depend on t E(Xt+k Xt ) m2 depends on k only (but does not depend on t) Representation of a Second Order Stationary Process A 2nd order stationary process can be represented : in the time domain by its autocorrelation function in the frequency domain by its power density spectrum = repartition of the energy of the signal on the different frequencies
Trafc Models S. Vaton Heavy-Tailed Distributions Long-Range Dependence Self Similarity Agregation Models Impact on Performance
Spectral Representation
Autocorrelation Function The autocorrelation function of a 2nd order stationary process is dened as follows : (k) = E(Xt+k Xt ) m2 Spectrum The power density spectrum of the process is the discrete Fourier transform of the autocorrelation function :
+
S () =
k =
(k) exp(ik)
The discrete time Fourier transform can be obtained as the output of the Fast Fourier Transform (FFT) algorithm.
Trafc Models S. Vaton Heavy-Tailed Distributions Long-Range Dependence Self Similarity Agregation Models Impact on Performance
LRD : denition
Long-Range Dependence (LRD) can be dened as follows :
slow decay of the autocorrelation function (k ) of a 2nd order stationary signal divergence in = 0 of the power density spectrum S () both denitions are equivalent
(k) = O(k2H 2 ), 0.5 < H < 1.0 S () = O(12H ), 0.5 < H < 1.0. Hurst parameter H
H is the Hurst parameter of the process the process is LRD if 0.5 < H 1 the closest to 1 is H , the strongest the LRD phenomenon is in the case of Markovian models, autoregressive models, etc... the Hurst parameter is H = 0.5 (no LRD) trafc analyses nd out most of the time H 0.80
Trafc Models S. Vaton Heavy-Tailed Distributions Long-Range Dependence Self Similarity Agregation Models Impact on Performance
0.6
0.4
0.2
-0.2
Trafc Models S. Vaton Heavy-Tailed Distributions Long-Range Dependence Self Similarity Agregation Models Impact on Performance
1e+10
1e+09
1e+09 1e+08
S ()
1e+08
S ()
LAN Traffic
0.001
1e+07
1e+07
10000
10000
AR2
0.001
1000 0.0001
0.01
0.1
1000 0.0001
0.01
0.1
Trafc Models S. Vaton Heavy-Tailed Distributions Long-Range Dependence Self Similarity Agregation Models Impact on Performance
LRD : signatures
Signatures of LRD :
slow decay of the autocorrelation function divergence of the power density spectrum in = 0 bad averaging properties : when the signal is averaged, strong variations remain in the averaged signal variance-time analysis is based on these bad averaging properties Variance-Time Analysis if Xt is a LRD process with Hurst parameter H then var( 1 m Xi )
i=1,m m
O(m2H 2 )
on the contrary, in the case of classical models (MAP, etc...) 1 1 ) it holds that var( m i=1,m Xi ) = O(m
Trafc Models S. Vaton Heavy-Tailed Distributions Long-Range Dependence Self Similarity Agregation Models Impact on Performance
Variance-Time Plot : average the signal on windows of m samples compute the variance sm of the averaged signal repeat for different values of m plot sm against m in log-log scale : straight line with slope (2H 2)
Trafc Models S. Vaton Heavy-Tailed Distributions Long-Range Dependence Self Similarity Agregation Models Impact on Performance
measured trafc trace (measurements over a LAN) synthetic trafc (AR(2) model)
10000 "AR2" "LAN Traffic"
1000
sm
100
10
1 1 10
100
1000
Trafc Models S. Vaton Heavy-Tailed Distributions Long-Range Dependence Self Similarity Agregation Models Impact on Performance
Fractional Gaussian Noise (fGn) popular particular case of LRD process dened by the following equation (1 B )d Xt = t where
d is a fractional coefcient : 0 < d < 0.5 k k k (1 B )d = k=0 Cd (1) B where B is the shift operator i.e. BXt = Xt1 t x 1 (d+1) k Cd = (k+1)( dt dk+1) where (x) = 0 e t
t W GN (0, 2 )
XH (t)
XH (t)
H=0.5
3 2
H=0.65
-1
-1
-2
-2
-3
-3
-4 0 200 400
t
H=0.8
600
800
1000
-4 0 200 400
600
800
1000
H=0.95
XH (t)
XH (t) t
-1
-1
-2
-2
-3
-3
-4 0 200 400
600
800
1000
Trafc Models S. Vaton Heavy-Tailed Distributions Long-Range Dependence Self Similarity Agregation Models Impact on Performance
Autocorrelation function
H
0.8 0.6 0.1 0.4 1
LRD
0.01
slope 2H2
0.2
-0.2
LRD
0 20 40
-0.4
60
80
60
80
100
Trafc Models S. Vaton Heavy-Tailed Distributions Long-Range Dependence Self Similarity Agregation Models Impact on Performance
Trafc Models S. Vaton Heavy-Tailed Distributions Long-Range Dependence Self Similarity Agregation Models Impact on Performance
Generation of a fGN
Method based on the FFT let Xt be a fGn, and S () its power density spectrum Xt is a 2nd order stationary process and thus Xt is harmonizable Xt =
2 it 1/2 ()dB () 0 e S
IN converges to Xt (in the sense of the L2 norm) Fast Fourier Transform algorithm (FFT)
n S 1/2 ( 2N ) FFT 2 N n ,
N 1 n=0 e
it 2n N
n S 1/2 ( 2N )
0nN 1
2 N n ,
1tN
Trafc Models S. Vaton Heavy-Tailed Distributions Long-Range Dependence Self Similarity Agregation Models Impact on Performance
Self-similarity : Principle
Intuitive idea : deterministic self-similarity
deterministic self-similarity : the same pattern is repeated on all scales statistical self-similarity : same characteristics (bursts, etc...) at all time scales
Trafc Models S. Vaton Heavy-Tailed Distributions Long-Range Dependence Self Similarity Agregation Models Impact on Performance
Self-Similarity in Trafc
Paxson and Floyd, The failure of Poisson modelling self-similarity of the cumulated workload variations are present at all time scales no typical time scale for bursts, etc... the cumulated workload is highly variable Self-Similarity denition of self-similarity signatures of self-similarity relation with LRD example : fractional Brownian motion
Trafc Models S. Vaton Heavy-Tailed Distributions Long-Range Dependence Self Similarity Agregation Models Impact on Performance
Denition of Self-Similarity
Statistical self-similarity : denition Let Yt be a continuous-time process. Yt is self-similar, with self-similarity coefcient H , if {Yat , t 0} = {aH Yt , t 0} equality between the laws of the processes a change of time scale does not change the shape of the self-similar process the interesting case is the case when 0.5 < H < 1.0
(d)
Trafc Models S. Vaton Heavy-Tailed Distributions Long-Range Dependence Self Similarity Agregation Models Impact on Performance
plot IDC (t) against t in log-log scale ; straight line with slope (2H 1) the case of the Poisson process corresponds to H = 0.5
Trafc Models S. Vaton Heavy-Tailed Distributions Long-Range Dependence Self Similarity Agregation Models Impact on Performance
S,
S (t, k )
k 1 avec
t,k )2 X
i=1,k
Xt+i
Trafc Models S. Vaton Heavy-Tailed Distributions Long-Range Dependence Self Similarity Agregation Models Impact on Performance
R/S index (Pox diagram) R/S index : results If Yt is self-similar with Hurst parameter H then E(R(t, k)/S (t, k)) = O(kH ) Pox diagram : plot the values of R/S against k in log-log scale ; straight line with slope H
if H is in the range 0.5 < H < 1.0 then the process is self-similar ; otherwise straight line with slope 0.5
Trafc Models S. Vaton Heavy-Tailed Distributions Long-Range Dependence Self Similarity Agregation Models Impact on Performance
a self-similar process is never 2nd order stationary (contrary to LRD processes) Which model should we choose ?
if the time series looks stationary then a LRD process can be used cumulated workload : self-similar process ; example : fractional Brownian motion, stable Levy motion, etc... workload (trafc per second, per minute, etc...) : LRD process ; example : fractional Gaussian noise, fARIMA, etc...
Trafc Models S. Vaton Heavy-Tailed Distributions Long-Range Dependence Self Similarity Agregation Models Impact on Performance
Refresher : standard Brownian motion Bt continuous time Gaussian process stationary and independent increments Fractional Brownian motion Bt stationary increments distribution of the increments (H ) (H ) Bt Bs N (0, 2 |t s|2H ) distribution of the increments Bt Bs N (0, 2 |t s|)
(H )
(fBm)
Trafc Models S. Vaton Heavy-Tailed Distributions Long-Range Dependence Self Similarity Agregation Models Impact on Performance
, t 0}
the case of the standard Brownian motion is the case H = 0.5 Fractional Brownian motion and fractional Gaussian noise let Xt be the increment process Xt associated to the (H ) fractional Brownian motion Bt Xt = Bt
(H )
Bt1
(H )
Trafc Models S. Vaton Heavy-Tailed Distributions Long-Range Dependence Self Similarity Agregation Models Impact on Performance
Agregation Models : some temptative explanations of some phenomenons observed in trafc (self-similarity of the cumulated workload, LRD in the series of trafc/minute, interarrival times, etc...) principle : the superposition of short-term memory effects at the microscopic level can result in some long-memory effects at the macroscopic level most popular model :
superposition of a large number of ON/OFF sources with Pareto distributed ON and/or OFF periods... Willinger, Taqqu et al., Self-Similarity Through High Variability : Statistical Analysis of LAN Trafc Data, 1997
Trafc Models S. Vaton Heavy-Tailed Distributions Long-Range Dependence Self Similarity Agregation Models Impact on Performance
Trafc Models S. Vaton Heavy-Tailed Distributions Long-Range Dependence Self Similarity Agregation Models Impact on Performance
Trafc Models S. Vaton Heavy-Tailed Distributions Long-Range Dependence Self Similarity Agregation Models Impact on Performance
Impact of LRD impact of LRD in trafc on some parameters of QoS (delay, losses, . . . ) Markovian models of trafc
are dramatically optimistic since they undervalue the probability of some rare events (buffer overows, very long delays, etc...) if the trafc is in fact LRD
Trafc Models S. Vaton Heavy-Tailed Distributions Long-Range Dependence Self Similarity Agregation Models Impact on Performance
Summary
Other Models
heavy-tailed distributions, Long Range Dependence, self-similarity signatures in trafc explanations of LRD in trafc impact on performance (delays)
Questions ?