1. Networked Media Lab.1 Weighted Backpressure Scheduling in IEEE 802.11 Wireless Mesh Networks Jaeyong Yoo jyyoo@nm.gist.ac.kr 23-11-09

2. Networked Media Lab.2 Background - Link / Packet Scheduling- N1 N2N5 N7N6 N4 Sender 1 Receiver 1 Sender 2 Receiver 2 Packet Scheduling –Which queue should be serviced first? Link Scheduling –Which link should be activated first? Objective –Throughput optimal –Fairness optimal –Stability Achievable Rate of S1 (Red Flow) Achievable Rate of S2 (Blue Flow) Objective 1. Throughput Optimality Objective 2. Fairness Optimality Objective 3. Stability

3. Networked Media Lab.3 Scheduling Research Map Notation: My queue length=X, upstream queue length=Y MWS (Maximum Weight Scheduling) a.k.a Back-pressure Scheduling GMS (Greedy Maximal Scheduling) Schedule policy: X - Y ’92 TAC ‘95 UCB Throughput efficiency of GMS ‘09 Mobihoc Distributed GMS ‘06 INFOCOM Capacity Region of GMS ‘08 INFOCOM Interference Condition GMS ‘08 INFOCOM Schedule policy: X (a.k.a no message passing) Tradeoff study between Message passing vs no message passing ‘08 Mobihoc Scheduling without “frequent” message passing ‘09 TWC Q-CSMA Yet Published Schedule policy: Y EZ-flow ‘09 CoNEXT Implemented System 2009 INFOCOM: DiffQ Implemented System 2008 Mobicom: Horizon Scheduling Policy Evolves Time Flies Schedule policy: βX + γY (WBS) Weighted Backpressure Scheduling

4. Networked Media Lab.4 Motivational Argument Queue non-stability while applying backpressure in IEEE 802.11 wireless mesh networks Query: Why non-stability comes? Despite the fact that many articles say backpressure is stable! –Previous implementation work DID NOT provide rigorous analysis on this part –Hence, we are doing this –Anyway, what is our major suspect?

5. Networked Media Lab.5 Suspect: Implementational approximation Network of Dream Theory to Practical Link-scheduling Assumption 1: Globally synchronized slotted access Link-scheduling Assumption 2: Perfect link schedule (At least fine-grained priority access) Link-scheduling Assumption 3: Immediate Link Schedule Many implicit assumption (Do not agree with reality) IEEE 802.11-based Wireless Mesh Networks Backpressure Link-scheduling Assumption 2: Quantized Priority Access (4 levels) Link-scheduling Assumption 2: Priority queue (Queuing delay) Approximation arrow Many other constraints (Even “currently unknown”)

6. Networked Media Lab.6 Main Argument Under the following conditions, –Quantized Priority Link Access, –Scheduling Time Delay (MAC layer queuing delay), and –Heterogeneous Link Qualities, Backpressure (X-Y) does not provide stable network queues But, β X + γ Y, with β < - γ –There exists values (β, γ, β < - γ) that stabilize network queues –Under the following conditions Quantization Level > 2 Any scheduling time delay “non-critical” link-quality heterogeneity, a.k.a drift

7. Networked Media Lab.7 System Model General n-hop case –Throughput model –Queue evolution model –Define drift Difference between two adjacent link’s throughput –Configurable queue limit: C, quantization step: L Abstract factor that contains link errors and other flow impact

8. Networked Media Lab.8 Main Result of 2-hop case In 2-hop case –Under the drift condition of “non-critical” drift –If beta and gamma satisfies below two conditions, the network queue becomes stable –With converging point to

9. Networked Media Lab.9 Let’s validate Experimental Validation Method –Implement the described system –By changing beta & gamma, observe the behavior

10. Networked Media Lab.10 System Implementation madwifi Priority queues Bypassing (Using PF_PACKET + RAW_SOCK + IPPROTO_RAW)Kernel Click f1f2 f3 f4f5 QQQQQ Per Flow table Sched P PPPP Choose Highest Schedule Priority athhal Antenna A view of a router below IP layer Scheduling Determination Taken Scheduling Action Taken Discrepancy of Scheduling time

11. Networked Media Lab.11 System Implementation (cont’d) Effort to minimize MAC-layer queuing delay

12. Networked Media Lab.12 System Implementation (cont’d) Notable bugs –Unordered packet delivery Unordered queue length monitor [fixed by filtering through ID field of IP header] –Madwifi 5 th queue problem 5 th queue has much higher access probability even if we change cwmin

13. Networked Media Lab.13 Experimentation Env. Experimentation time = 3 minutes Change beta and gamma (step 0.2) –0 <= Beta < 2 –-2 < Gamma <= 0 –Total 121 points C = 100 L = 30 Madwifi Priority queues: 8 queues Manipulate channel quality change by inserting random error –Drift (Up and Down drift)

14. Networked Media Lab.14 Beta, Gamma, Out of range Mis-configuration of Beta and Gamma (2, -1.8) (1.2, -1.0) (0.4, -0.2)

15. Networked Media Lab.15 Beta, Gamma, Out of range (Deep Inside) Inside of [0.4, -0.2], C=100 N1 RET N2 RET N1 MAC N2 MAC N2 Queue N1 monitored Queue

16. Networked Media Lab.16 Beta, Gamma, Out of range (Deep Inside) Inside of [0.4, -0.2], C=100 N2 Queue N1 monitored Queue Quantized Priority Space

17. Networked Media Lab.17 N1 RET N2 RET N1 MAC N2 MAC N2 Queue N91monitored Queue Beta, Gamma, In range (Deep Inside) Inside of [0.6, -0.2], C=100

18. Networked Media Lab.18 Beta, Gamma, In range (Deep Inside) Inside of [0.6, -0.2], C=100 N2 Queue N1 monitored Queue Quantized Priority Space Link-error impulse

19. Networked Media Lab.19 Beta, Gamma, In range (Deep Inside: Drift changing) N2 Queue N1 monitored Queue Quantized Priority Space Inside of [0.6, -0.2], C=100 Drift direction changing point

20. Networked Media Lab.20 Overall Comparison From Model (Dark point represents stable point) Gamma 0-2 Beta 0 2

21. Networked Media Lab.21 Overall Comparison (cont’d) Average Queue length (Drift Down) Average Queue length (Drift Up)

22. Networked Media Lab.22 Overall Comparison (cont’d) Deviation Queue length (Drift Down) Deviation Queue length (Drift Up)

23. Networked Media Lab.23 Conclusion Very first analysis of queue stability with considering real- world constraint –Three constraints Delay, Drift, Quantized Priority Space –Provides rule of thumb Next step –Analysis is focusing on “averaged behavior” –What about network variance? –Will narrow down the choice of beta and gamma Adaptive algorithm that finds beta and gamma

24. Networked Media Lab.24 Backup slides

25. Networked Media Lab.25 Scheduling Framework Node j Node j-1Node j+1 q j,1 q j,n q( q j,1, Q j+1,1 )Q j,1 p(q j,1, Q j+1,1 )P j,1 Q j+1, 1 Q j, 1 q j-1,1 q j-1,n q j+1,1 q j+1,n q j,i means queue length of node j of i flow P j,i means the priority of node j of i flow

26. Networked Media Lab.26 Positioning Under Scheduling Framework †. Two functions q, p can describe various invariants of packet scheduling 1. Back-pressure scheduling [ q(x, y) = x ] [p(x, y) = x – y ] 2. Ez-Flow [ q(x,y) = x] [p(x, y) = -y] 3. PNCP [ q(x,y) = (x+y)/2] [ p(x, y) = x – y ] 4. No-message passing [ no necessary of q] [ p(x, y) = x ] 5. Our proposal [q(x,y) = x + αy] [q(x, y) = βx + γy]

27. Networked Media Lab.27 Throughput Model Validation

28. Networked Media Lab.28 Physical Behavior with Three Conditions L 0t C t Priority Queue len Send Faster d d S R

29. Networked Media Lab.29 Imagine of slotted contention status - 1 10 9 8 7 6 5 4 3 2 1 10 9 8 7 6 5 4 3 2 1 0 Delay = 6 C=5 [C-1 = 5] [7-C = 2]  N1 wins [C-2 = 3] [8-C = 3]  Let’s say N1 wins [C-3 = 2] [7-C = 2]  Let’s say N1 wins [C-4 = 1] [8-C = 3]  Let’s say N2 wins [C-5 = 0] [7-C = 2]  Let’s say N2 wins [C-6 = -1] [6-C =1]  Let’s say N2 wins [C-7 = -2] [5-C =0]  Let’s say N2 wins [C-8 = -3] [4-C =0]  Let’s say N2 wins [C-7 = -2] [3-C =-2]  Let’s say N1 wins Going up at b=3

30. Networked Media Lab.30 10 9 8 7 6 5 4 3 2 1 10 9 8 7 6 5 4 3 2 1 0 Delay = 6 C=5 γ :0.5 [C-0.5 = 4.5] [7-2.5 = 4.5]  Let’s say N1 wins [C-1 = 4] [8-2.5 = 5.5]  N2 wins [C-1.5 = 3.5] [7-2.5 = 4.5]  N2 wins [C-2 = 3] [6-2.5 = 3.5]  N2 wins [C-2.5 = 2.5] [5-2.5 = 2.5]  Let’s say N1 wins [C-3 = 2] [4-2.5 = 1.5]  Let’s say N2 wins Going up at b=3 Imagine of slotted contention status - 2

31. Networked Media Lab.31 Relationship between delay & drift C-b(n-d) = b(n)-C –Increasing case b(n-d) = b(n) – d C-b(n) – d = b(n)-C b(n) = C – d/2 –Decreasing case b(n-d) = b(n) + d C-b(n) + d = b(n)-C b(n) = C + d/2 Conjecture Amplitude of oscillation will follow Ud