1 / 18

Minimum-Delay Load-Balancing Through Non-Parametric Regression F. Larroca and J.-L. Rougier

Minimum-Delay Load-Balancing Through Non-Parametric Regression F. Larroca and J.-L. Rougier. IFIP/TC6 Networking 2009 Aachen, Germany, 11-15 May 2009. Introduction. Current traffic is highly dynamic and unpredictable

xena
Download Presentation

Minimum-Delay Load-Balancing Through Non-Parametric Regression F. Larroca and J.-L. Rougier

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Minimum-Delay Load-Balancing Through Non-Parametric RegressionF. Larroca and J.-L. Rougier IFIP/TC6 Networking 2009 Aachen, Germany, 11-15 May 2009

  2. Introduction • Current traffic is highly dynamic and unpredictable • How may we define a routing scheme that performs well under these demanding conditions? • Possible Answer: Dynamic Load-Balancing • We connect each Origin-Destination (OD) pair with several pre-established paths • Traffic is distributed in order to optimize a certain function • Function fl (rl ) is typically a convex increasing function that diverges as rl → cl; e.g. mean queuing delay • Why queuing delay? Simplicity and versatility IFIP/TC6 Networking 2009 F. Larroca and J.-L. Rougier

  3. Introduction • A simple model (M/M/1) is always assumed • What happens when we are interested in actually minimizing the total delay? • Simple models are inadequate • We propose: • Make the minimum assumptions on fl (rl ) (e.g. monotone increasing) • Learn it from measurements instead • Optimize with this learnt function IFIP/TC6 Networking 2009 F. Larroca and J.-L. Rougier

  4. Agenda • Introduction • Attaining the optimum • Delay function approximation • Simulations • Conclusions IFIP/TC6 Networking 2009 F. Larroca and J.-L. Rougier

  5. Problem Definition • Queuing delay on link l is given by Dl(rl) • Our congestion measure: weighted mean end-to-end queuing delay • The problem: • Since fl (rl ):=rl Dl (rl ) is proportional to the queue size, we will use this value instead IFIP/TC6 Networking 2009 F. Larroca and J.-L. Rougier

  6. Congestion Routing Game • Path P has an associated cost fP : where fl(rl) is continuous, positive and non-decreasing • Each OD pair greedily adjusts its traffic distribution to minimize its total cost • Equilibrium: no OD pair may decrease its total cost by unilaterally changing its traffic distribution • It coincides with the minimum of: IFIP/TC6 Networking 2009 F. Larroca and J.-L. Rougier

  7. Congestion Routing Game • What happens if we use ? • The equilibrium coincides with the minimum of: • To solve our problem, we may play a Congestion Routing Game with • To converge to the Equilibrium we will use REPLEX • Important: fl(rl) should be continuous, positive and non-decreasing IFIP/TC6 Networking 2009 F. Larroca and J.-L. Rougier

  8. Agenda • Introduction • Attaining the optimum • Delay function approximation • Simulations • Conclusions IFIP/TC6 Networking 2009 F. Larroca and J.-L. Rougier

  9. Cost Function Approximation • What should be used as fl (rl)? • That represents reality as much as possible • Whose derivative (fl(rl)) is: • continuous • positive => fl (rl) non-decreasing • non-decreasing => fl (rl) convex • To address 1 we estimate fl (rl) from measurements • Convex Nonparametric Least-Squares (CNLS) is used to enforce 2.b and 2.c : • Given a set of measurements {(ri,Yi)}i=1,..,N find fNϵ F where F is the set of continuous, non-decreasing and convex functions IFIP/TC6 Networking 2009 F. Larroca and J.-L. Rougier

  10. Cost Function Approximation • The size of F complicates the problem • Consider instead G (subset of F) a family of piecewise-linear convex non-decreasing functions • The same optimum is obtained if we change F by G • We may now rewrite the problem as a standard QP one IFIP/TC6 Networking 2009 F. Larroca and J.-L. Rougier

  11. Cost Function Approximation • This regression function presents a problem: its derivative is not continuous (cf. 2.b) • A soft approximation of a piecewise linear function: • Our final approximation of the link-cost function: IFIP/TC6 Networking 2009 F. Larroca and J.-L. Rougier

  12. An Example IFIP/TC6 Networking 2009 F. Larroca and J.-L. Rougier

  13. Agenda • Introduction • Attaining the optimum • Delay function approximation • Simulations • Conclusions IFIP/TC6 Networking 2009 F. Larroca and J.-L. Rougier

  14. NS-2 simulations • The considered network: IFIP/TC6 Networking 2009 F. Larroca and J.-L. Rougier

  15. NS-2 simulations • Alternative (“wrong”) training set: IFIP/TC6 Networking 2009 F. Larroca and J.-L. Rougier

  16. Agenda • Introduction • Attaining the optimum • Delay function approximation • Simulations • Conclusions IFIP/TC6 Networking 2009 F. Larroca and J.-L. Rougier

  17. Conclusions and Future Work • We have presented a framework to converge to the actual minimum total mean delay demand vector • Two shortcomings of our framework: • fl(rl) is constant outside the support of the observations • Links with little or no queue size have a negligible cost • Possible Solution: Add a “patch” function that is negligible with respect to fl(rl) except at high loads • How does fl(rl) behaves over time? Does it change? How often? • How does our framework performs when compared with other mechanisms or simpler models? • Faster and/or more robust alternative regression methods? • Is REPLEX the best choice? IFIP/TC6 Networking 2009 F. Larroca and J.-L. Rougier

  18. Thankyou! Questions? IFIP/TC6 Networking 2009 F. Larroca and J.-L. Rougier

More Related