1 / 23

ITMANET PI Meeting September 2009

ITMANET Nequ-IT. Focus Talk (PI Neely): Reducing Delay in MANETS via Queue Engineering. ITMANET PI Meeting September 2009. Queueing Theory 101:. Slotted Queueing System Random Packet Arrivals rate λ ( packets/slot ) Random Service Opportunities rate μ (packets/slot)

farica
Télécharger la présentation

ITMANET PI Meeting September 2009

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. ITMANET Nequ-IT Focus Talk (PI Neely): Reducing Delay in MANETS via Queue Engineering ITMANET PI Meeting September 2009

  2. Queueing Theory 101: • Slotted Queueing System • Random Packet Arrivals rateλ (packets/slot) • Random Service Opportunities rate μ (packets/slot) If:ε = μ – λ = proximity to boundary of capacity Then: Average Delay = O(1/ε) [Note: O(1/ε) tradeoff holds only for stochastic arrivals and/or channels] Example: Bernoulli Arrivals and Service μ λ E{Delay} ε μ 1- λ E{Delay} = = O(1/ε) λ μ - λ

  3. Stochastic Network Optimization Theory 101: T/R T/R T/R T/R T/R T/R T/R • Random Packet Arrivals, Random Channels, MANET • Unknown Traffic, Channel Probabilities, Mobility Model • “Backpressure + Max-Weight + Flow Control” result from greedy action to minimize “drift-plus-penalty” *[Neely 03, 06]: *Minimize:Δ(Q(t)) + (1/ε)Ε{Penalty(t)|Q(t)} [ε = a positive parameter chosen as desired, Δ(Q(t)) = “Quadratic Lyapunov Drift”] E{Delay} ε max utility utility

  4. Stochastic Network Optimization Theory 101: T/R T/R T/R T/R T/R T/R T/R Theorem[PI Neely: MIT thesis 2003, F&T text 2006]: Under the drift-plus-penalty algorithm with any desired ε>0: Distance to Optimal Utility < O(ε) Average end-to-end delay < O(1/ε) Holds for: • General Performance Objectives (thruput, thruput-utility, energy) • General Multi-Hop MANETS, Any size, General ergodic mobility E{Delay} ε max utility utility

  5. Stochastic Network Optimization Theory 101: T/R T/R T/R T/R T/R T/R T/R Theorem[PI Neely: MIT thesis 2003, F&T text 2006]: Under the drift-plus-penalty algorithm with any desired ε>0: Distance to Optimal Utility < O(ε) Average end-to-end delay < O(1/ε) Holds for: • General Performance Objectives (thruput, thruput-utility, energy) • General Multi-Hop MANETS, Any size, General ergodic mobility Is this the optimal delay tradeoff??? E{Delay} ε max utility utility

  6. Optimal Network Delay Tradeoff Theory: O(1/ε) is NOT the optimal delay tradeoff! Depending on the network situation, for single-hop nets, we know the optimaldelay tradeoff is either: • Square Root Law: Average Delay > Ω(sqrt[1/ε]) • Logarithm Law: Average Delay > Ω(log[1/ε]) These Results were proven by Nequ-IT PIs: • PI Berry [Information Theory 2002] • Single Queue System with Energy Optimization • Known Traffic and Channel Statistics • PI Neely [JSAC 2006, Information Theory 2007] • Multi-Queue System with Energy or Thruput-Utility Optimization • Unknown Traffic and Channel Statistics • Different control technique. Holds in single-hop, limited multi-hop (not as general as drift-plus-penalty)

  7. Re-Visit the Drift-Plus-Penalty Algorithm: Drift-Plus-Penalty (Quadratic Lyapunov Algorithm): • Disadvantages: Only gives the (sub-optimal) [O(ε), Ο(1/ε)] tradeoff • Advantages: Works in more extensive (multi-hop, mobile) networks Observations: • Algorithm uses Queue Backlog to inform the stochastic optimization • Queue Backlogs must go high to get good utility performance • Information in Relative Magnitudes of Backlogs, and in the Oscillations Idea: Use “Fake Backlog” to trick the optimizer! Two “Magic Numbers”: • M2: Hard to compute • M1: Easy to compute

  8. Re-Visit the Drift-Plus-Penalty Algorithm: Drift-Plus-Penalty (Quadratic Lyapunov Algorithm): • Disadvantages: Only gives the (sub-optimal) [O(ε), Ο(1/ε)] tradeoff • Advantages: Works in more extensive (multi-hop, mobile) networks Observations: • Algorithm uses Queue Backlog to inform the stochastic optimization • Queue Backlogs must go high to get good utility performance • Information in Relative Magnitudes of Backlogs, and in the Oscillations Idea: Use “Fake Backlog” to trick the optimizer! Two “Magic Numbers”: • M2: Hard to compute • M1: Easy to compute Actual backlog under M1 M1 place-holder backlog M1

  9. Re-Visit the Drift-Plus-Penalty Algorithm: Drift-Plus-Penalty (Quadratic Lyapunov Algorithm): • Disadvantages: Only gives the (sub-optimal) [O(ε), Ο(1/ε)] tradeoff • Advantages: Works in more extensive (multi-hop, mobile) networks Observations: • Algorithm uses Queue Backlog to inform the stochastic optimization • Queue Backlogs must go high to get good utility performance • Information in Relative Magnitudes of Backlogs, and in the Oscillations Idea: Use “Fake Backlog” to trick the optimizer! Two “Magic Numbers”: • M2: Hard to compute • M1: Easy to compute M2 place-holder backlog M2

  10. New Result 1: Magic Number M1 [Neely, Asilomar, Dec. 08] Advantages of Magic Number M1: • Can be computed easily • Works for any MANET • Improves delay with no loss of utility! • 30% Delay Savings in example Limitations: • Biggest M1 savings for min-penalty problems (e.g., energy minimization) • Only a constant factor delay reduction, still have [O(ε), Ο(1/ε)] tradeoff Example MANET: Uses diversity backpressure routing (DIVBAR) w/o place-holders Avg. Power Avg. Backlog with place-holders 1/ε (where 1/ε = V) 1/ε (where 1/ε = V)

  11. New Result 1: Magic Number M1 [Neely, Asilomar, Dec. 08] Advantages of Magic Number M1: • Can be computed easily • Works for any MANET • Improves delay with no loss of utility! • 30% Delay Savings in example Limitations: • Biggest M1 savings for min-penalty problems (e.g., energy minimization) • Only a constant factor delay reduction, still have [O(ε), Ο(1/ε)] tradeoff Example MANET: Uses diversity backpressure routing (DIVBAR) w/o place-holders Avg. Power Avg. Backlog with place-holders 1/ε (where 1/ε = V) 1/ε (where 1/ε = V)

  12. New Result 1: Magic Number M1 [Neely, Asilomar, Dec. 08] Advantages of Magic Number M1: • Can be computed easily • Works for any MANET • Improves delay with no loss of utility! • 30% Delay Savings in example Limitations: • Biggest M1 savings for min-penalty problems (e.g., energy minimization) • Only a constant factor delay reduction, still have [O(ε), Ο(1/ε)] tradeoff Example MANET: Uses diversity backpressure routing (DIVBAR) w/o place-holders Avg. Power Avg. Backlog with place-holders 1/ε (where 1/ε = V) 1/ε (where 1/ε = V)

  13. New Result 1: Magic Number M1 [Neely, Asilomar, Dec. 08] Advantages of Magic Number M1: • Can be computed easily • Works for any MANET • Improves delay with no loss of utility! • 30% Delay Savings in example Limitations: • Biggest M1 savings for min-penalty problems (e.g., energy minimization) • Only a constant factor delay reduction, still have [O(ε), Ο(1/ε)] tradeoff Example MANET: Uses diversity backpressure routing (DIVBAR) w/o place-holders Avg. Power Avg. Backlog with place-holders 1/ε (where 1/ε = V) 1/ε (where 1/ε = V)

  14. New Result 1: Magic Number M1 [Neely, Asilomar, Dec. 08] Advantages of Magic Number M1: • Can be computed easily • Works for any MANET • Improves delay with no loss of utility! • 30% Delay Savings in example Limitations: • Biggest M1 savings for min-penalty problems (e.g., energy minimization) • Only a constant factor delay reduction, still have [O(ε), Ο(1/ε)] tradeoff Example MANET: Uses diversity backpressure routing (DIVBAR) w/o place-holders Avg. Power Avg. Backlog with place-holders 1/ε (where 1/ε = V) 1/ε (where 1/ε = V)

  15. New Result 1: Magic Number M1 [Neely, Asilomar, Dec. 08] Advantages of Magic Number M1: • Can be computed easily • Works for any MANET • Improves delay with no loss of utility! • 30% Delay Savings in example Limitations: • Biggest M1 savings for min-penalty problems (e.g., energy minimization) • Only a constant factor delay reduction, still have [O(ε), Ο(1/ε)] tradeoff Example MANET: Uses diversity backpressure routing (DIVBAR) w/o place-holders Avg. Power Avg. Backlog with place-holders 1/ε (where 1/ε = V) 1/ε (where 1/ε = V)

  16. New Result 1: Magic Number M1 [Neely, Asilomar, Dec. 08] Advantages of Magic Number M1: • Can be computed easily • Works for any MANET • Improves delay with no loss of utility! • 30% Delay Savings in example Limitations: • Biggest M1 savings for min-penalty problems (e.g., energy minimization) • Only a constant factor delay reduction, still have [O(ε), Ο(1/ε)] tradeoff Example MANET: Uses diversity backpressure routing (DIVBAR) w/o place-holders Avg. Power Avg. Backlog with place-holders 1/ε (where 1/ε = V) 1/ε (where 1/ε = V)

  17. New Result 1: Magic Number M1 [Neely, Asilomar, Dec. 08] Advantages of Magic Number M1: • Can be computed easily • Works for any MANET • Improves delay with no loss of utility! • 30% Delay Savings in example Limitations: • Biggest M1 savings for min-penalty problems (e.g., energy minimization) • Only a constant factor delay reduction, still have [O(ε), Ο(1/ε)] tradeoff Example MANET: Uses diversity backpressure routing (DIVBAR) w/o place-holders Avg. Power Avg. Backlog with place-holders 1/ε (where 1/ε = V) 1/ε (where 1/ε = V)

  18. New Result 1: Magic Number M1 [Neely, Asilomar, Dec. 08] Advantages of Magic Number M1: • Can be computed easily • Works for any MANET • Improves delay with no loss of utility! • 30% Delay Savings in example Limitations: • Biggest M1 savings for min-penalty problems (e.g., energy minimization) • Only a constant factor delay reduction, still have [O(ε), Ο(1/ε)] tradeoff Example MANET: Uses diversity backpressure routing (DIVBAR) w/o place-holders Avg. Power Avg. Backlog with place-holders 1/ε (where 1/ε = V) 1/ε (where 1/ε = V)

  19. New Result 1: Magic Number M1 [Neely, Asilomar, Dec. 08] Advantages of Magic Number M1: • Can be computed easily • Works for any MANET • Improves delay with no loss of utility! • 30% Delay Savings in example Limitations: • Biggest M1 savings for min-penalty problems (e.g., energy minimization) • Only a constant factor delay reduction, still have [O(ε), Ο(1/ε)] tradeoff Example MANET: Uses diversity backpressure routing (DIVBAR) w/o place-holders Avg. Power Avg. Backlog with place-holders 1/ε (where 1/ε = V) 1/ε (where 1/ε = V)

  20. New Result 2: Magic Number M2 [Huang, Neely, WiOpt 2009] Lagrange Multiplier M2 Result of Huang-Neely WiOpt 09: • Steady state probability distribution for queue backlog decays exponentiallyabout a suitably defined “Lagrange Multiplier” of a corresponding non-stochastic problem. • Works for the drift-plus-penalty algorithm [Neely 2003, 2006] • Significantly tightens the prior result on proximity to Lagrange multiplier by Eryilmaz-Srikant 06 (they used a “fluid-limit” argument)

  21. New Result 2: Magic Number M2 [Huang, Neely, WiOpt 2009] Lagrange Multiplier M2 Advantages of Magic Number M2: • Dramatically improves delay. Backlog “rarely” falls below M2 • Achieves an improved delay tradeoff: [O(ε), O(log2[1/ε])] Within a log-factor of achieving the optimal log() delay tradeoff! Limitations: • Harder to compute M2 (ideally should know the “Lagrange Multiplier”) • Works for single-hop and limited classes of multi-hop • Must drop a small fraction of packets (O(ε)) to compensate when cross M2.

  22. Concluding Remarks: Experimental Work at USC This analysis also motivates and fundamentally explains recent USC experimental results showing dramatic delay improvement for backpressure by: Moeller, Sridharan, Krishnamachari,Gnawali, “Backpressure Routing Made Practical,” Submitted to Hotnets 09. See also tech report at: http://anrg.usc.edu/www/index.phpPublications_by_Year#techreport2009 Experimental Results next slide

  23. Concluding Remarks: Experimental Work at USC • 40 Node Tiny OS2.x Multi-Hop Sensor Network • Moeller et. al. develop 2 simplified implementations of “effective” M2 algorithm without computing M2!! (one answer: “Use Last-In-First-Out” ) • Dramatic Backpressure Delay Improvement (75-98%), for all but 1% of packets! • 50% improvement in throughput compared to conventional shortest path algs!

More Related