IEEEACM TRANSACTIONS ON NETWORKING VOL 15 NO 3

IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 15, NO. 3, JUNE 2007 Fairness and Load Balancing in Wireless LANs Using Association Control Yigal Bejerano, Member, IEEE, Seung-Jae Han, Member, IEEE, and Li (Erran) Li, Member, IEEE Presented by 范姜竣韋(C. W. Fan-Chiang) 許宴毅(Y. Y. Hsu)

Contents 2020/11/3 １ INTRODUCTION ２ SYSTEM DESCRIPTION ３ FAIRNESS AND LOAD BALANCING ４ ASSOCIATION CONTROL ALGORITHMS ５ ONLINE INTEGRAL-ASSOCIATION ６ SIMULATION RESULTS ７ CONCLUSION 2

Contents 2020/11/3 １ INTRODUCTION ２ SYSTEM DESCRIPTION ３ FAIRNESS AND LOAD BALANCING ４ ASSOCIATION CONTROL ALGORITHMS ５ ONLINE INTEGRAL-ASSOCIATION ６ SIMULATION RESULTS ７ CONCLUSION 3

Introduction v Load imbalance Each user associate itself with the AP that has the strongest RSSI(received signal strength indicator), while ignoring its load condition Some APs may idle v Solution Association control： • Balance the load via intelligently selecting the user. AP association 2020/11/3 4

Introduction(cont’d) v Association control can be used to achieve different objectives. To maximize the overall system throughput by shifting not from the fairness viewpoint More desirable goal： provide fair bandwidth allocation, while maximizing the minimal fair share of each user v In this paper, we present efficient algorithms Ensure max-min fair bandwidth allocation This goal obtained by balancing the load on the APs 2020/11/3 5

Contents 2020/11/3 １ INTRODUCTION ２ SYSTEM DESCRIPTION ３ FAIRNESS AND LOAD BALANCING ４ ASSOCIATION CONTROL ALGORITHMS ５ ONLINE INTEGRAL-ASSOCIATION ６ SIMULATION RESULTS ７ CONCLUSION 6

The Network Model Assume that adjacent APs use noninterfering channels Consider long-term fairness Greedy users that always have traffic consume all the allocated bandwidth 2020/11/3 7

System Discription 1. System requires relevant information on each user, such as and 2. It needs an algorithm to determine the user. AP association 3. It need a mechanism to enforce these association decisions the collected information is reported to a network operation center(NOC) • 2020/11/3 Periodically recalculates the optimal user association by using the offline algorithms 8

Periodic Offline Optimization v Motivation： By showing the weakness of the existing heuristic load balancing mechanisms Example： (b)：Least-loaded-first(LLF) Bandwidth：{4/3, 1, 4/3} b/4+b/2=1 b=4/3 (c)、(d)：strongest-signal-first(SSF) Case 1 Bandwidth：{8/7, 8/7} Case 2 Bandwidth：{8/3, 2} 2020/11/3 9

Wireless and Wired Bottlenecks Wireless link is generally considered as the bottleneck this assumption is not always valid Case I： fair user association only from the wireless perspective wireless: 0. 5 Mb/s to each user T 1 line: AP a: 0. 5 Mb/s to user 5, 6 Wireless link is the bottleneck AP b: 3/8 to its associated user Wired link is the bottleneck Case II： A fair user association 0. 5 Mb/s to each user over the wired and wireless channels 2020/11/3 10

Contents 2020/11/3 １ INTRODUCTION ２ SYSTEM DESCRIPTION ３ FAIRNESS AND LOAD BALANCING ４ ASSOCIATION CONTROL ALGORITHMS ５ ONLINE INTEGRAL-ASSOCIATION ６ SIMULATION RESULTS ７ CONCLUSION 11

FAIRNESS AND LOAD BALANCING v Two association models 1. Single-association model =integral-association 2. Multiple-association model =fractional-association Its bandwidth allocation is the aggregated bandwidth Denote by all the users that associated with AP a A denotes the set of APs that u U is associated with. 2020/11/3 12

A. Max-Min Fairness A bandwidth allocation is a matrix ： aggregated bandwidth to user Normalized bandwidth(NB)： NBV： sorted in increasing order AP a is required to serve user u a period of over the wireless channel and over the infrastructure link v If a bandwidth allocation is feasible if and v Fair service：all users have the same allocated bandwidth the degree of fairness may cause reduction of the throughput provide max-min fairness v v v 2020/11/3 13

A. Max-Min Fairness(cont’d) v max-min fairness No way to increase the bandwidth of a user without decreasing the bandwidth of another user with the same or less normalized bandwidth B 1={1, 1, 2, 2, 3} B 2={1, 1, 2} v a user association is termed max-min fair if its corresponding bandwidth allocation is max-min fair 2020/11/3 14

Example (b): ‧a feasible fair association ‧every user receive b=1 (c): ‧NBV (d): ‧NBV 1 ={1, 1, 1, 2, 2} ={1, 4/3, 4/3} 4/3 v NBV of a fractional max-min fairness allocation always ＞ ＝ NBV of the integral max-min fairness allocation v The users can be divided into fairness groups consists of all users that experience the same NB allocation 2020/11/3 15

B. Min-Max Load Balancing v The notion of load is not well defined Neither # of users associated with an AP nor its throughput reflect the AP’s load v the load of an AP needs to reflect its inability to satisfy the requirements of its associated users and as such it should be inversely proportional to the average bandwidth that they experience v We are able to extend existing load balancing techniques to balance the AP loads and obtain a fair service 2020/11/3 16

B. Min-Max Load Balancing (cont’d) v A fractional association is a matrix for each u U, holds. v Each specifies the fractional association of user u with AP a. Reflects the fraction of user u’s total flow that it expects to get from AP a v A fractional association is feasible if v User produces a load of on the wireless channel of AP and a load of on its backhaul link 2020/11/3 17

B. Min-Max Load Balancing (cont’d) v We define the load induced by user on AP to be the time that is required of AP to provide user a traffic volume of size 2020/11/3 18

B. Min-Max Load Balancing (cont’d) v Define the load vector of an association matrix sorted in decreasing order v 2020/11/3 19

Example 1 1 ¼+¼=1/2 2 2 ½+½=1 1 1 ¼+¼+¼=3/4 ¼+½=3/4 4/3 4/3 (c)：(integral) • 1 = {1, 1, 1/2} (d)：(fractional) • 3/4 = {1, 3/4} v APs can be partitioned into load groups contains all the APs with the same load assigned in decreasing order. 2020/11/3 20

B. Min-Max Load Balancing (cont’d) v normalize =1 1 2020/11/3 21

B. Min-Max Load Balancing (cont’d) v v In the following we refer to the load group of the most loaded APs and the corresponding fairness group as the bottleneck groups 2020/11/3 22

B. Min-Max Load Balancing (cont’d) v v Unfortunately, Theorem 5 is not satisfied in the case of a single association v Example: ={1, 1, 1/2} ={1, 1, 1, 2, 2} 2020/11/3 ={1, 1, 2} 23

Contents １ INTRODUCTION ２ SYSTEM DESCRIPTION ３ FAIRNESS AND LOAD BALANCING ４ ASSOCIATION CONTROL ALGORITHMS ５ ONLINE INTEGRAL-ASSOCIATION ６ SIMULATION RESULTS ７ CONCLUSION

ASSOCIATION CONTROL ALGORITHMS v A extension of the scheduling unrelated parallel machines problem For any ε< ½, there’s NO polynomial-time ( 1+ε ) approximation algorithm exists, unless P = NP v To seek for a complete min-max load balanced association. A 2 -approximation algorithm for unweighted users, A 3 -approximation algorithm for weighted users An optimal solution for fair time allocation.

A. ρ*-Approximation With Threshold v Intuitively, we would like to guarantee to each user a bandwidth of at least 1/ ρ of the bandwidth that it receives in the optimal integral solution, for a constant ρ≧ 1. v However, there is neither upper nor lower constant bounds for the ratio.

A. ρ*-Approximation With Threshold (cont’d)

A. ρ*-Approximation With Threshold (cont’d) v Our practical goal is to reduce the load of highly loaded APs, there is no need to balance the load of APs with load below a certain threshold T, where T is the maximal load that a user may generate on an AP as formulated in Recall that T is indeed a very small value and in practical 802. 11 networks T ≦ 1 s/Mb.

A. ρ*-Approximation With Threshold (cont’d) v v

B. Scheme Overview v Integral Load Balancing Algorithm.

1) Fractional Load Balancing Algorithm

B. Scheme Overview (cont’d) v we utilize a linear program, denoted as LP 1 which calculates a feasible association and also minimizes the maximal load on all the APs over both their wireless and wired channels v

B. Scheme Overview (cont’d) v

B. Scheme Overview (cont’d) v Bottleneck-group Detection Routine

2) The Rounding Method v

2) The Rounding Method a　　　b　　　c 1 1 X X 2 X 1 X 3 X 1/2 4 X ½ ½ 5 X X 1 Qa, 1={1} Qb, 1={4, 3} Qb, 2={3, 2} Qb, 3={2} Qc, 1={5, 4} Qc, 2={5} Ua a { 1 } Ub b { 4, 3, 2 1/2 Uc 1/2 } 1/2 { 1/2 4, 5 }

C. Analysis of the Unweighted Case v

C. Analysis of the Unweighted Case (cont’d) v

C. Analysis of the Unweighted Case (cont’d) v

D. Weighted Users v v v

D. Weighted Users (cont’d) v v v

D. Weighted Users (cont’d) v

E. Time Fairness v

E. Time Fairness (cont’d) 1/3 1/2

Contents １ INTRODUCTION ２ SYSTEM DESCRIPTION ３ FAIRNESS AND LOAD BALANCING ４ ASSOCIATION CONTROL ALGORITHMS ５ ONLINE INTEGRAL-ASSOCIATION ６ SIMULATION RESULTS ７ CONCLUSION

V. ONLINE INTEGRAL-ASSOCIATION v v

Contents １ INTRODUCTION ２ SYSTEM DESCRIPTION ３ FAIRNESS AND LOAD BALANCING ４ ASSOCIATION CONTROL ALGORITHMS ５ ONLINE INTEGRAL-ASSOCIATION ６ SIMULATION RESULTS ７ CONCLUSION

VI. SIMULATION RESULTS v Assume Distance 50 m 80 m 120 m 150 m bit rate 11 Mb/s 5. 5 Mb/s 2 Mb/s 1 Mb/s Maximum transmission range: 150 m 20 APs are located on 5 by 4 grid Number of users is either 100 or 250 Users are randomly positioned in a circleshaped hot spot with 150 m near the center of the simulation network Results are obtained from averaging 100 runs

VI. SIMULATION RESULTS (cont’d) v

VI. SIMULATION RESULTS (cont’d) v

VI. SIMULATION (cont’d) Each time. RESULTS slot replace 20% of users v The offline algorithm is invoked Every 15 time slots or when bottleneck difference exceeds 25%

Contents １ INTRODUCTION ２ SYSTEM DESCRIPTION ３ FAIRNESS AND LOAD BALANCING ４ ASSOCIATION CONTROL ALGORITHMS ５ ONLINE INTEGRAL-ASSOCIATION ６ SIMULATION RESULTS ７ CONCLUSION

VII. CONCLUSION v The problem of providing fair service to users and balancing the load among APs. This goal is achieved by intelligently determining the user-AP association. v Our simulations confirm that the proposed methods, indeed, achieve close to optimal load balancing and max-min fair bandwidth allocation, and significantly outperform popular heuristics. v Moreover, we show that in some cases, by balancing the load on the APs the overall network throughput is increased. In the future, we intend to develop a practical management system based on theoretical foundation presented in this study.