Department of Electronic Engineering (EE) City University of Hong Kong Performance Evaluation of Long Range Dependent Queues Chen Jiongze Supervisor: Moshe Zukerman Co-Supervisor: Ronald G. Addie Supported by Grants [CityU 124709] and [CityU 8/CRF/13G] Introduction Poisson Lomax Burst Process (PLBP) Queue Fractional Brownian motion (fBm) Queue Conclusions 2 Introduction Poisson Lomax Burst Process (PLBP) Queue Fractional Brownian motion (fBm) Queue Conclusions 3 Not enough capacity Angry customer Credit: http://inwritefield.com/2012/06/22/are-you-being-served/ 4 Too much capacity Lose Money Credit: http://www.neverpaintagain.co.uk/blog/how-to-lose-money-with-bad-home-improvement-choices/ 5 Capacity QoS Traffic Engineering Network Engineering Network Planning Credit: http://ineed.coffee/ 6 Data Data Data A Link Traffic Data Data … Data Data Data … … Credit: http://www.fastcodesign.com/1662881/infographic-of-the-day-the-facebook-map-of-the-world 7 Data Data Data A Link Traffic Data Data … Data Quality of Service (QoS) ? Data Data … … It is unrealistic to replicate the entire traffic on an Internet link! 8 A Link Traffic Internet Link Data … Data Modelling Queueing System Input process Queue 9 Input process Traffic Data Data Sampling Data … … A good traffic model • capture the nature of the traffic: Long Range Dependence (LRD); • A small number of parameters; • Amenable to analysis. Fitting the parameters Traffic model 10 [1] W. E. Leland, M. S. Taqqu, W. Willinger, and D. V. Wilson, “On the self-similar nature of ethernet traffic (extended version),” IEEE/ACM Trans. Networking, vol. 2, no. 1, pp. 1–15, Feb. 1994. 11 A process, X={Xt, t=1,2,…}, with mean m and variance σ2. Autocovariance function, γ(k) = E[(Xt-m)(Xt+k-m)], decays slower than exponential. IID – Independent and Identically Distributed MMPP – Markov Modulated Poisson Process IID Poisson 0 LRD MMPP k 12 A process, X={Xt, t=1,2,…}, with mean m and variance σ2. Autocovariance function, γ(k) = E[(Xt-m)(Xt+k-m)], decays slowly. Autocorrelation function (ACF), ρ(k) = γ(k)/σ2, follows Hurst parameter H : the measure of the degree of the LRD. 0.5 < H < 1 the process is LRD The aggregate process of X with interval t, X(t) follows 13 Input process Traffic Data Data Sampling Data … … Important statistics of traffic: mean (m), variance (σ2) and Hurst parameter (H). Fitting the parameters Traffic model 14 LRD process Input traffic process Single Server Queue (SSQ) Link Overflow probability QoS LRD Queue LRD process Mean (m) Variance (σ2) Hurst parameter (H) SSQ Output P(Q>x)? SSQ with ∞ buffer Steady state Queue Size (Q) Service rate (μ) 15 Hurst parameter Capacity Variance Hurst parameter Variance Blocking probability Buffer size Buffer size Mean Blocking probability Helpful for Traffic Engineering Mean Capacity Network Engineering & Network Planning 16 Introduction Poisson Lomax Burst Process (PLBP) Queue Fractional Brownian motion (fBm) Queue Conclusions 17 The process is based on a stream of bursts. Burst arrivals following a Poisson process with rate λ [bursts/s]. Burst durations, d, are i.i.d. Lomax random variables with parameters γ and δ. Each burst contributes work with constant rate r [B/s]. PLBP: {At, t≥0}, where At is the work contributed by all bursts during the interval (0,t]. 18 Exponential () Lomax (, ) … … t Work 4r 3r 2r r … t PLBP 19 Exponential () Common probability mass function (PMF) G … … t Bt 4 3 2 1 … t M/G/∞ process 20 Exponential () Pareto (, ) … … t Work 4r 3r 2r r … t Poisson Pareto Burst Process (PPBP) 21 The complementary cumulative distribution function (CCDF) of • Pareto distribution • Lomax distribution d: burst duration; γ: shape parameter; δ: scale parameter. Advantages of Lomax: • takes care of small bursts; • δ is no more minimum value of burst duration. 22 Mean (m(t)): Variance (σ2(t)): 1<γ<2LRD 23 Input PLBP with parameters: λ, γ, δ and r. SSQ Output SSQ with ∞ buffer and service rate, μ. Obtain the overflow probability, P(Q>x), by: • Analytical result: the Quasi-stationary (QS) approximation, • Simulation: the fast simulation method. 24 τ PLBP Process Short burst process (Sτ) Long burst process (Lτ) t t+τ t t+τ For a certain period τ, the probability of Q>x is Number of long bursts, η, is Poisson distributed with mean as λE(d)P(ω>τ). λE(d): the mean number of the existing burst at t; ω: the forward recurrence time of a Lomax RV. QS: the steady state queue size of an SSQ fed by Sτ . Assuming Sτ is Gaussian, we can obtain P(QS>x) by [2]. QS approximation: [2] R. G. Addie, P. Mannersalo, and I. Norros, Performance formulae for queues with Gaussian input, ser. Teletraffic Engineering in a Competitive World. Elsevier Science, Jun. 1999, pp. 1169–1178. 25 Conventional simulation method … Fast simulation method Initial short bursts Initial long burst T 26 27 Introduction Poisson Lomax Burst Process (PLBP) Queue Fractional Brownian motion (fBm) Queue Conclusions 28 Three characteristic features: A continuous Gaussian process; Self-similar with parameter H; Stationary increment. A normalized fBm process, B={B(t), t≥0}: B(0) = 0, E[B(t)] = 0 for all t≥0; Var[B(t)] = t2H for all t≥0; 0.5<H<1 LRD; Covariance function: 29 The increment process – fractional Gaussian Noise (fGn): Cumulative work arrival process: X = {X(t), t≥0}, where X(t) = mt+σB(t), thus E[X(t)] = mt, Var[X(t)] = σ2t2H; the mean and variance of its increment process are m and σ2. Input fBm process with parameters: m, σ and H. SSQ Output SSQ with ∞ buffer and service rate, μ. 30 The mean net input ι = m – μ, so an fBm queue has three parameters, ι, σ and H. By Reich’s formula [3], the queue size (Q) is Exact solution for P(Q>x) for H = 0.5 by Harrison [4]: No exact results for P(Q>x) for H ≠ 0.5. [3] E. Reich, “On the integrodifferential equation of takács I,” Annals of Mathematical Statistics, vol. 29, pp. 563–570, 1958. [4] J. M. Harrison, Brownian motion and stochastic flow systems. New York: John Wiley and Sons, 1985. 31 By Norros [5]: It holds in sense that [5] I. Norros, “A storage model with self-similar input,” Queueing Syst., vol. 16, no. 3-4, pp. 387–396, Sep. 1994. 32 By Hüsler and Piterbarg [6]: where C is a certain constant and the it holds in sense that No method to determine C. Since for RHS ∞ as x 0. [6] J. Hüsler and V. Piterbarg, “Extremes of a certain class of Gaussian processes,” Stochastic Processes and their Applications, vol. 83, no. 2, 33 pp. 257 – 271, Oct. 1999. Revise Hüsler and Piterbarg’s approach by supposing that it is not the CCDF, but rather the density whose character remains stable for x near ∞. where the density cf(x) is characterized according to 34 For x near ∞: Dominant We have 35 Let and , we have where and Γ denotes the Gamma function. It holds in the sense that 36 Our approximation vs. asymptotics by Hüsler and Piterbarg Advantages: ▪ a distribution ▪ accurate for full range of buffer size ▪ provides ways to derive c Disadvantages: ▪ Slightly less accurate for very large x 37 Discrete-time simulation. Divide time into N intervals of equal length Δt. Qn denotes the queue size at the end of nth interval, defined by Lindley’s equation: where Q0=0, each interval. is the amount of work arriving in Difficulty: Discrete time Δt 0 Continuous time 38 For a given H, simulate for different Δt with one sequence of standard fGn, , with mean ι and variance v1. A new sequence is defined by where s(Δt) and m(Δt) are chosen so that has the appropriate mean and variance. 39 40 41 42 43 The density function of Q: The density function of Amoroso Distribution: = g = 0, d = β, p = ν and a = α−1/ν 44 45 Negative arrivals Appropriate for 1) large buffer size; 2) σ is large relative to m. Gaussian Appropriate for high multiplexing. To illustrate the weaknesses, we compared it with the PPBP model. 46 Small buffer size Large buffer size CISCO routers 47 Deriving the inverse function of our approximation, we have the dimensioning formula as where μ*: capacity; m: mean of the input process; σ2: variance of the input process; H: Hurst parameter; ε: required overflow probability; q: buffer threshold; G-1(): inverse regularised incomplete Gamma function. 48 49 50 Introduction Poisson Lomax Burst Process (PLBP) Queue Fractional Brownian motion (fBm) Queue Conclusions 51 Main contributions: PLBP queues The PLBP model (a variant of PPBP) is proposed. An approximation based on the QS algorithm is provided. A fast simulation method is applied. fBm queues New results for queueing performance and link dimensioning are derived. Important statistics of fBm queues are provided. An efficient approach for simulation is proposed. The weaknesses of the fBm model are discussed. 52 53