1. Distributed Association Control in Shared Wireless Networks Krishna C. Garikipati and Kang G. Shin University of Michigan-Ann Arbor

2. Shared Wireless Networks  Modes of operation  Advantages Improves network coverage and capacity Under-utilized APs put to use Peer-to-peer sharingPublic sharing

3. Key Features  Uncoordinated Access Points Ad-hoc deployment No global policy  Backhaul Limited Wireless capacity > wired capacity  Throughput Inefficiency RSSI based AP selection Unfairness + low bandwidth utilization Internet User AP ADSL

4. Association Control  An important problem 1 Control of user associations to prevent overloading and/or starvation of users AB AB B A C B A C Throughput Crucial for the success of sharing 1 “Seven Ways that HetNets are a Cellular Paradigm Shift”, IEEE Communications Magazine, March 2013

5. Setup  Variables Set of users,  Throughput Set of APs, Association of user is Association vector, where Set of users connected to AP is Backhaul capacity MAC overhead MCS Rate Airtime fraction Equal for all users connected to same AP

6. Association Control Problem  Balancing throughput via user associations where is defined as the proportional fair utility  How to solve it without a central controller ? Utility Maximization NP-hard => intractable for large search space

7. Related Work None of them achieve PF in a distributed way  Utility based approaches WorkFairnessCoordination  [A. Kumar and V. Kumar 05] Optimal association of stations and APs  [Bejarano et al. 03] Load-balancing of APs  [Li et al. 08] Approx. algo. for Multi-Rate WLANs Centralizedmax-min  [Kauffmann et al. 07] Self Organization of WLANs proportional delay proportional Centralized Distributed Centralized

8. This Work  Feasibility of association control without global coordination  Optimal randomized solution with probabilistic associations  Sub-optimal greedy approach with performance bounds Dense networks: Backhaul limited: Concept of Marginal utility Steady state distribution:

9. Randomized Approach

10.  User associates with APs probabilistically  Desired steady state distribution Lemma : For every, is an increasing function in. Moreover, as, Connects for a random duration, scans and switches Generated Markov Chain: where is a fixed parameter

11. Update Process  Poisson clock  Discretization Users have i.i.d clocks with inter-tick duration Scan is triggered at a clock tick Equivalent DTMC is where is the global poisson clock T1T1 time User update process T2T2 T3T3 T4T4 Scanning Association

12. Update Process, e.g.,  Gibbs sampler Association prob. of user at a clock tick Markov Chain is aperiodic, irreducible is the steady state distribution Not distributed as user requires global information to compute One-step transition probability is

13. Distributed Update Process  Objective function separation where utility of AP is defined as  Define Marginal Utility for each AP w.r.t user where is set of users connected to AP except

14. Distributed Update Process  New Update rule

15. Distributed Update Process  New Update rule User can obtain locally through scanning Current Association Probing AP

16. Distributed Update Process  New Update rule User can obtain locally through scanning Current Association Probing AP

17. Distributed Update Process  New Update rule User makes a decision on switching Current Association Selects next association with prob. distribution

18. Distributed Update Process  New Update rule Completely distributed and asynchronous User initiates reassociation with selected AP Old Association New Association

19. Partial Information  Marginal utility from subset of APs is known Due to partial scanning or probe frame losses Probability of knowing utility from AP is Current Association Probing AP

20. Partial Information  Marginal utility from subset of APs is known Due to partial scanning or probe frame losses Probability of knowing utility from AP is Theorem 1 The generated Markov chain has steady state distribution where

21. Partial Information  Marginal utility from subset of APs is known Due to partial scanning or probe frame losses Probability of knowing utility from AP is Theorem 1 The generated Markov chain has steady state distribution where Theorem 2 The expected utility in steady state satisfies where and

22. Greedy Approach

23. Best Association  User associates in a deterministic way Greedy approach to randomization At clock tick, user chooses AP Theorem 3 The Best Association converges almost surely. Every optimal association is an equilibrium association. Results in Nash Equilibrium which satisfies the property for all and all

24. Best Association  User associates in a deterministic way Greedy approach to randomization At clock tick, user chooses AP Theorem 3 The Best Association converges almost surely. Every optimal association is an equilibrium association. Results in Nash Equilibrium which satisfies the property for all and all Equilibrium state is not easy to find

25. Best Association  Two scenarios Dense (collocated) NetworkBackhaul limited Users connect to same set of APs and at same PHY rate All APs are backhaul limited and wireless settings are irrelevant

26. Dense Networks  User index can be dropped Number of users associated with each AP, Theorem 4 Every equilibrium association is globally optimal, that is Utility of AP where, are constants Theorem 5 It takes at most N re-associations to reach equilibrium; each user switches at most once Concave

27. Backhaul limited  Wireless parameters can be ignored Number of users associated with each AP, Theorem 6 Every equilibrium association satisfies the lower bound, Each user has different neighborhood Concave Utility of AP, assume

28. Simulation

29.  Performance in random topology Greedy approach converges to almost optimal solution Association control performs significantly better than RSSI approach Partial scanning leads to slower convergence

30. Simulation  Comparison with other distributed policies Best Association gives the highest fairness Slight reduction in throughput due to PF fairness

31. Conclusion  Association control in shared WLANs Greedy heuristic performs close to optimal Achievable using a distributed mechanism  Extendable to Heterogeneous Networks ?

32. Thank you Krishna C. Garikipati gkchai@eecs.umich.edu