41
Link Dimensioning for Fractional Brownian Input Chen Jiongze PhD student, Electronic Engineering Department, City University of Hong Kong Supported by Grant [CityU 124709] Moshe Zukerman Electronic Engineering Department,, City University of Hong Kong Ron G. Addie Department of Mathematics and Computing, University of Southern Queensland, Australia

Link Dimensioning for Fractional Brownian Input Chen Jiongze PhD student, Electronic Engineering Department, City University of Hong Kong Supported by

Embed Size (px)

Citation preview

Link Dimensioning for Fractional Brownian

Input

Chen JiongzePhD student,

ElectronicEngineering Department,City University of Hong

Kong

Supported by Grant [CityU 124709]

Moshe ZukermanElectronic

Engineering Department,,City University of Hong Kong

Ron G. AddieDepartment of Mathematics

and Computing,University of

Southern Queensland, Australia

Outline:

• Background• A New Analytical Result of an FBM Queue• Simulation• Link Dimensioning• Discussion & Conclusion

Outline:

• Background• A New Analytical Result of an FBM Queue• Simulation• Link Dimensioning• Discussion & Conclusion

Fractional Brownian Motion (fBm)• process of parameter H, Mt

H are as follows:• Gaussian process N(0,t2H)• Covariance function:

• For H > ½ the process exhibits long range dependence

traffic Queue

traffic Queue

Long Range Dependence

Gaussian By

Central limittheorem

• Its statistics match those of real traffic (for example, auto-covariance function)

- Gaussian process & LRD• A small number of parameters

- Hurst parameter (H), variance• Amenable to analysis

Is fBm a good model? YES

Outline:

• Background• A New Analytical Result of an FBM Queue• Simulation• Link Dimensioning• Discussion & Conclusion

A New Analytical Result of an fBm Queue

traffic Queue

Queuing Model

fBm trafficHurst parameter (H)variance (σ1

2)drift / mean rate of traffic (λ)

Single server Queue with ∞ buffersservice rate (τ)steady state queue size (Q)

mean net input (μ = λ - τ)

Analytical results of (fBm) Queue

No exact results for P(Q>x) for H ≠ 0.5

Existing asymptotes:•By Norros [9]

[9] I. Norros, “A storage model with self-similar input,” Queueing Syst., vol. 16, no. 3-4, pp. 387–396, Sep. 1994.

Analytical results of (fBm) Queue

Existing asymptotes (cont.):•By Husler and Piterbarg [14]

[14] J. H¨usler and V. Piterbarg, “Extremes of a certain class of Gaussian processes,” Stochastic Processes and their Applications, vol. 83, no.

2, pp. 257 – 271, Oct. 1999.

Approximation of [14] is more accurate for large x but with no way provided to calculate •Our approximation:

Analytical results of (fBm) Queue

• Our approximation VS asymptote of [14]:

•Advantages:• a distribution• accurate for full range of u/x• provides ways to derive c

•Disadvantages:• Less accurate for large x (negligible)

Analytical results of (fBm) Queue

[14] J. H¨usler and V. Piterbarg, “Extremes of a certain class of Gaussian processes,” Stochastic Processes and their Applications, vol. 83, no.

2, pp. 257 – 271, Oct. 1999.

Outline:

• Background• A New Analytical Result of an FBM Queue• Simulation• Link Dimensioning• Discussion & Conclusion

Simulation

• Basic algorithm (Lindley’s equation):

n 0 1 2 …

Un (Mb) 1.234 – 0.3551 0.743 …

m(t) (Mb) -0.5 -0.5 -0.5 …

Qn (Mb) 0 max(0, 1.234 – 0.5)=0.734

max (0, 0.734 – 0.3551 – 0.5)=0

1 ms

Length of Un = 222 for different Δt, it is time-consuming to generate Un for different time unit)

An efficient approachInstead of generating a new sequence of numbers, we change the “units” of work (y-axis).

1ms -> Δt ms

variance of the fBn sequence (Un): V

Rescale the Y-asix

An efficient approachInstead of generating a new sequence of numbers, we change the “units” of work (y-axis).

1 unit = S instead of 1 where

Rescale m and P(Q>x)•m = μΔt/S units, so

•P(Q>x) is changed to P(Q>x/S)

Only need one fBn sequence

Simulation Results• Validate simulation

Simulation Results

Simulation Results

Simulation Results

Simulation Results

Outline:

• Background• A New Analytical Result of an FBM Queue• Simulation• Link Dimensioning• Discussion & Conclusion

Link Dimensioning• We can drive dimensioning formula by

Incomplete Gamma function:

Gamma function:

Finally

Link Dimensioning

where C is the capacity, so .

Link Dimensioning

Link Dimensioning

Link Dimensioning

Link Dimensioning

Link Dimensioning

Outline:

• Background• Analytical results of a fractional Brownian motion (fBm)

Queue• Existing approximations• Our approximation

• Simulation• An efficient approach to simulation fBm queue• Results

• Link Dimensioning

• Discussion & Conclusion

Discussion

• fBm model is not universally appropriate to Internet traffic• negative arrivals (μ = λ – τ)

• Further work• re-interpret fBm model to

• alleviate such problem• A wider range of parameters

ConclusionIn this presentation, weIn this presentation, we•considered a queue fed by fBm input

•derived new results for queueing performance and link dimensioning

•described an efficient approach for simulation

•presented • agreement between the analytical and the simulation results

• comparison between our formula and existing asymptotes

• numerical results for link dimensioning for a range of examples

References:

References:

References:

Q & AQ & A

Background• Self-similar (Long Range Dependency)

• “Aggregating streams of traffic typically intensifies the self similarity (“burstiness”) instead of smoothing it.”[1]

• Very different from conventional telephone traffic model(for example, Poisson or Poisson-related models)

• Using Hurst parameter (H) as a measure of “burstiness”

[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.

Background• Self-similar (Long Range Dependence)

• “Aggregating streams of traffic typically intensifies the self similarity (“burstyiness”) instead of smoothing it.”[1]

• Very different from conventional telephone traffic model(for example, Poisson or Poisson-related models)

• Using Hurst parameter (H) as a measure of “burstiness”• Gaussian (normal) distribution

• When umber of source increases

[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.

[6] M. Zukerman, T. D. Neame, and R. G. Addie, “Internet traffic modeling and future technology implications,” in Proc. IEEE INFOCOM

2003,vol. 1, Apr. 2003, pp. 587–596.

process of Real traffic Gaussian process [2]Central limit

theorem

Especially for core and metropolitan Internet links, etc.

Analytical results of (fBm) Queue

• A single server queue fed by an fBm input process with- Hurst parameter (H)- variance (σ1

2)

- drift / mean rate of traffic (λ)- service rate (τ)- mean net input (μ = λ - τ)- steady state queue size (Q)

• Complementary distribution of Q, denoted as P(Q>x), for H = 0.5:

[16]

[16] J. M. Harrison, Brownian motion and stochastic flow systems. New York: John Wiley and Sons, 1985.