1. Rice University Origins and Solutions of Proportional Unfairness in IEEE 802.11 Multi-hop Wireless Networks Jingpu Shi PhD Thesis Defence Department of Electrical and Computer Engineering Rice University November, 2007

2. November 2007ECE Department, Rice University Jingpu Shi 2 Motivation Gaining popularity. e.g., mesh networks in Philadelphia, San Francisco, Corpus Christi... IEEE 802.11 widely deployed as the MAC. However, IEEE 802.11 incurs unfairness problem. Multi-hop network: not all nodes hear each other's transmissions.

3. November 2007ECE Department, Rice University Jingpu Shi 3 Thesis Objective To investigate unfairness in 802.11 multi-hop networks and propose solutions to improve fairness. Specifically, to –demonstrate existence of unfairness. –analyze the origins of unfairness. –analytically model the origins of unfairness. –propose solutions to improve fairness. –evaluate the proposed solutions.

4. November 2007ECE Department, Rice University Jingpu Shi 4 Fairness Background There are many definitions for fairness, we consider proportional fairness. Proportional fairness. –Proposed by Kelly in 1997. –Definition: A vector of rates x is proportionally fair if it is feasible and if for any other feasible vector x*, the aggregate of proportional changes is zero or negative: Example: x2 x1 x2 + 3x1 = 0 p1 p2 p3 –It is a maximizer of the strictly concave log utility function:

5. November 2007ECE Department, Rice University Jingpu Shi 5 Outline Analysis and modelling of unfair MAC contention. Analysis and modelling of joint effect of MAC and congestion control on unfairness. Solutions to improve fairness. Conclusion.

6. November 2007ECE Department, Rice University Jingpu Shi 6 Prior work [Bianchi 00] [Kumar 05 ] [Sharma 06] etc. Single clique, same channel view This work Different channel view, Topological factors. [Bharghavan 94] [Kanodia 02] etc. some unfair scenarios, but not modelled. Analytically modelled. Objective: understand and model unfair MAC contention in multi-hop 802.11 networks where nodes have incomplete channel information. Unfair MAC Contention

7. November 2007ECE Department, Rice University Jingpu Shi 7 SNR > CS thresold Basic Two-flow Four-node Layout A ba B AB Ab aB ab Senders A, B receivers a, b. A sender must be connected to its respective receiver. Dotted line: two nodes are connected (SNR is above carrier sense threshold.) Nodes from one flow may hear nodes from the other flows. 12 possible combinations.

8. November 2007ECE Department, Rice University Jingpu Shi 8 12 Possible Scenarios, 3 Categories Asymmetric Incomplete state (AIS) Sender Connected (SC) Symmetric Incomplete State (SIS)

9. November 2007ECE Department, Rice University Jingpu Shi 9 Analysis of AIS Assume there is no capture effect. The two flows have different view of the channel. And experience different collision probabilities. t A a Bb RCDATAARC ARC ARC A Bb's view: medium is idle. R Aa's view: medium is busy. RRR

10. November 2007ECE Department, Rice University Jingpu Shi 10 Modeling AIS The channel “ private view ” of a node: … … SuccessfulIdle Collision t Busy state Modeled as a renewal-reward process. Throughput (pkt/s) = P [event Ts occurs] Average duration of event Compute probability of each state. Generality is validated in [Garetto Infocom 06].

11. November 2007ECE Department, Rice University Jingpu Shi 11 Model Validation: Simulation Set Up SimulatorNs2 Ver. 2.1b7a Data rate11Mbps Basic rate2Mbps Data sizeVarying SIFS, DIFS, EIFS10,50,364 us Time slot20 us PLCP length192 bits @ 1 Mbps (CWmin, CWmax)(31,1023) Short, Long retry limit7,4

12. November 2007ECE Department, Rice University Jingpu Shi 12 Validation – Model vs. Simulation 0 200 400 600 800 1000 200400600800100012001400 Packet Throughput (pkt/s) Data Payload Size (bytes) 0 200 400 600 800 1000 200400600800100012001400 ns - Flow B model - Flow B ns - Flow A model - Flow A With RTS/CTSWithout RTS/CTS ns - Flow A model - Flow A ns - Flow B model - Flow B

13. November 2007ECE Department, Rice University Jingpu Shi 13 Validation – Model vs. Experiments 0 200 400 600 800 1000 Packet Throughput (pkt/s) 0 200 400 600 800 1000 With RTS/CTSWithout RTS/CTS Flow B Flow A model TFA Flow B Flow A model TFA TFA network, SMC 2532-b 802.11b 200 mW power, 1500 Bytes, 11Mbps

14. November 2007ECE Department, Rice University Jingpu Shi 14 Analysis of SIS Long-term fair. –Due to symmetric topology. Short-term unfair. –Once a flow loses, it is harder and harder for it to win back, until it resets its contention status. A a B b

15. November 2007ECE Department, Rice University Jingpu Shi 15 Modeling SIS The two flows closely coupled. Propose a two-dimensional Markov model with the state being { backoff stage of A, backoff stage of B }. Computing the transition probability is the key. Two key steps: –Simplifying assumption to remove memory property: replace uniform distribution with geometric distribution. –Handle hidden terminals.

16. November 2007ECE Department, Rice University Jingpu Shi 16 SIS: Model vs. Simulation CaseThroughput (pkt/s) Collision probability Time scale of unfairness (ms) RTS/CTS 7 stages 218 216 0.25 235 223 RTS/CTS 9 stages 229 230 0.11 0.09 982 1156 Basic access 4 stages 125 107 0.69 0.75 15 Basic access 7 stages 222 220 0.37 0.38 59 60 ns model System’s bi-stability: With large probability, the system is in one of the two stable states: (1) Flow Aa captures the channel. (2) Flow Bb captures the channel.

17. November 2007ECE Department, Rice University Jingpu Shi 17 Summary of analysis of unfair MAC contention. –Systematically and analytically studied two-flow scenarios. –Nodes can contend unfairly under IEEE 802.11 due to incomplete channel information. –The AIS class incurs long-term unfairness. –The SIS class incurs short-term unfairness.

18. November 2007ECE Department, Rice University Jingpu Shi 18 Objective: to understand what is the joint effect of TCP and MAC on fairness. Prior Work: –[Gerla Infocom 05, Sivakumar MobiHoc 03] TCP poor performance due to TCP's large congestion window. Solution: TCP should not inject more than what can be served. E.g. limit or fix TCP window to a small value. This thesis: –Will show that window flow control itself (even with very small window) on top of 802.11 can lead to unfairness. –Analytically model the unfairness scenario. TCP and MAC Joint Effect

19. November 2007ECE Department, Rice University Jingpu Shi 19 Basic Scenario AB GW Two upload TCP flows (Many variations discussed in the thesis). The basic topology. –(A,B) (B,GW) are in range, A and GW are hidden terminals to each other. –Important topology in mesh networks. AB GW Two-hop TCP One-hop TCP Abstraction, not necessarily a real line topology.

20. November 2007ECE Department, Rice University Jingpu Shi 20 Outline of Basic Scenario Analysis Analyze the case with RTS/CTS on. –Same analysis applies to RTS/CTS off. Analyze the MAC behaviour first, then add window flow control. Mathematically model the scenario. Propose and evaluate a solution.

21. November 2007ECE Department, Rice University Jingpu Shi 21 MAC Bi-stability DATA Aggregate ACK A B GW GW is also contending. Consider A's and GW's traffic first, add B's traffic later. Variation of the SIS class. Bi-stable state: either A transmits and GW is in high backoff, or GW transmits and A is in high backoff. Success state and fail state. CW=2CW min CW=2 2 CW min CW=2CW min CW=CW min CW=2 2 CW min CW=2 k CW min CW=CW min A's RTS GW's RTS A's RTS GW's RTS

22. November 2007ECE Department, Rice University Jingpu Shi 22 MAC Bi-stability Take B into consideration: B is in range of both A and GW, so B's packets interleave with A's and GW's packets. Either pair (A,B) transmit and GW is in high backoff, or pair (B,GW) transmit and A is in high backoff. ABGW TCP ACK TCP DATA A trafficGW traffic B traffic Multiple packet burst (GW,B)Multiple packet burst (A,B)

23. November 2007ECE Department, Rice University Jingpu Shi 23 Impact of TCP on Bi-stability Two nested transport loops, the inner and the outer loop. GWBA DATA ACK GWBA ACK DATA Inner loop Outer loop GWBA ACK DATA Transport loops change the duration of bi-stability. (A, B) burst, GW in fail state, the burst size is limited by: –TCP window size. (GW, B) burst, A in fail state, self-sustaining loop: –TCP ACK are generated.

24. November 2007ECE Department, Rice University Jingpu Shi 24 Two-hop Node’s Severe MAC and TCP Penalty ABGW TCP DATA TCP ACK CW=2CW min CW=2 2 CW min CW=2CW min CW=CW min CW=2 2 CW min A's MAC penalty A B A Pkt loss Timeout A's TCP penalty

25. November 2007ECE Department, Rice University Jingpu Shi 25 Modelling the Basic Topology Only model sliding window mechanism –Will show sliding window along can induce unfairness. Assume geometric backoff counter distribution. Six-dimensional Markov Model –Eight channel states: 1 idle, 1 collision, 6 success transmission states. –Possible switching times: mini-slot boundaries. –System state: { Q 1,Q 2,Q 3,Q g,W a,W g } where Q g = Q 4 +Q 5, W a and W g are minimum contention window of A and GW, respectively.

26. November 2007ECE Department, Rice University Jingpu Shi 26 Compute Transition Probability and Solve the Model { Q 1,Q 2,Q 3,Q g,W a,W g } { Q 1 -1,Q 2 +1,Q 3,Q g,0,W g } The transition probability is : where e is the success probability f is the transmission duration of node A in time slots. A example: when a success transmission state occurs on link 1:

27. November 2007ECE Department, Rice University Jingpu Shi 27 Solution to Improve Fairness Unfairness is caused by collision between A and GW. To improve fairness –We need to decrease the steady state probability of those system states where Q1 > 0 and Qg > 0 –This can be done by increasing CWmin at node B. AB GW Q1Q1 QgQg

28. November 2007ECE Department, Rice University Jingpu Shi 28 Reduced Backlog at A and GW With Increased CWmin at B AB GW Indeed, the probability that both A and GW are backlogged is reduced. Default CWmin Default CWmin Increase CWmin Q1Q1 QgQg

29. November 2007ECE Department, Rice University Jingpu Shi 29 Model Validation Validate the model in three levels: How well does it agree with ns to validate the geometric distribution assumption? Can the model capture the trend of throughput changes when CWmin at node B increases in real networks ? Can the solution motivated by the model improve fairness in real networks ?

30. November 2007ECE Department, Rice University Jingpu Shi 30 Default CWmin Varying CWmin Default CWmin Model Vs. Simulation ns-2.28, 11M bps, RTS/CTS on, 1500 bytes, TCP window fixed. Model agrees with simulation results. ABGW

31. November 2007ECE Department, Rice University Jingpu Shi 31 Model Vs. Real Experiment Model captures the trend of throughput changes when CWmin increases. Motivates solution: to increase CWmin of GW’s one-hop neighbor. Mirror Mesh Linux Operating System (kernel 2.6). Atheros wireless card (Madwifi v.0.9.2 driver). Data rate fixed to 11Mbps. RTS/CTS on.

32. November 2007ECE Department, Rice University Jingpu Shi 32 Two or More Branches Same analysis for basic scenario applies to two or more branches –Two hop flow(s) contend unfairly with one-hop flow(s) Solution: to increase CWmin of GW’s one-hop neighbors AB GW TCP ACK TCP DATA C TCP ACK

33. November 2007ECE Department, Rice University Jingpu Shi 33 Solution Validation: Mirror Mesh Setup GW B A

34. November 2007ECE Department, Rice University Jingpu Shi 34 Validation For Basic Topology With RTS/CTS On Increase CWmin ABGW CWmin at 1st hop nodes Severe unfairness with default CWmin, log utility = -3.8 Improved fairness with larger CWmin at 1 st hop nodes, log utility = -1.23 Maximum log utility = 3.2

35. November 2007ECE Department, Rice University Jingpu Shi 35 Validation For Two Branches RTS/CTS ON Increase CWmin B->GW A->GW C->GW CWmin at 1 st hop nodes Severe unfairness with default CWmin, log utility = -3.8. Improved fairness with larger CWmin at 1 st hop nodes, log utility = -1.23. Maximum log utility = 3.2.

36. November 2007ECE Department, Rice University Jingpu Shi 36 1 st hop CWmin = 128 Validation for Long Hop Chain RTS/CTS ON 21GW 34 Increase CWmin NS simulation. Severe unfairness with default CWmin, log utility = -11.9763. Improved fairness with larger CWmin, log utility = -6.1721. Maximum log utility = 3.6931.

37. November 2007ECE Department, Rice University Jingpu Shi 37 Summary –Analysis for the basic topology. Identified the two-hop topology where unfairness manifests. Analyzed the joint effect of MAC and window flow control on unfairness Modelled the basic topology. –Solution: minimum contention window policy. Increase the CWmin of 1 st hop node(s).

38. November 2007ECE Department, Rice University Jingpu Shi 38 Utilizing multiple channels to improve fairness Asynchronous Multi-channel Coordination Protocol (AMCP)

39. November 2007ECE Department, Rice University Jingpu Shi 39 Motivation and Challenges Multiple channels add capacity to the network. Multiple channels provide opportunity to improve throughput of the flows that would receive little throughput due to unfair contention in single channel case. Challenges: –Single-channel unfair contention. –Multi-channel coordination problems. Multi-channel hidden terminal problem. Missing receiver problem.

40. November 2007ECE Department, Rice University Jingpu Shi 40 AMCP General Description Asynchronous protocol, one radio per node. One common control channel, multiple data channels. Reserve common channel and data channel differently. Nodes only contend for channels clear of traffic. Self-learning channel hopping.

41. November 2007ECE Department, Rice University Jingpu Shi 41 Related Work SSCH [ Bahl MobiCom04], MMAC [So MobiHoc 04] All existing protocols are designed to improve aggregate throughput.

42. November 2007ECE Department, Rice University Jingpu Shi 42 Thesis Contributions –Identified topologies where unfair contention manifests. –Explained how MAC independently or jointly with window flow control induce unfairness, when nodes don't have complete channel information. –Proposed analytical models to study unfairness in multi-hop networks. A model that allows different nodes to have different channel view. A two-dimensional Markov model to capture the coupling between two flows. A six-dimensional Markov model to capture the joint effect of MAC and window flow control on fairness. –Designed protocol solutions to improve fairness. A simple minimum contention window policy for mesh networks. AMCP to utilize multiple channels.

43. November 2007ECE Department, Rice University Jingpu Shi 43 Publications 1.M. Garetto, J. Shi and E. Knightly, "Modeling Media Access in Embedded Two- Flow Topologies of Multi-hop Wireless Networks," in Proceedings of ACM MobiCom'05, Cologne, Germany, August 2005. 2.J. Shi, T. Salonidis and E. Knightly, "Starvation Mitigation Through Multi-Channel Coordination in CSMA-based Wireless Networks," in Proceedings of ACM MobiHoc'06, Florence, Italy, May 2006. 3.J. Shi, O. Gurewitz, V. Mancuso, J. Camp and E. Knightly, "Measurement and Modeling of the Origins of Starvation in Congestion Controlled Mesh Networks," Proceedings of Infocom'08, Phoenix, AZ, April, 2008.

44. November 2007ECE Department, Rice University Jingpu Shi 44 Thanks !