1. MDG: Measurement-Driven Guidelines for 802.11 WLAN Design Ioannis Broustis, Konstantina Papagiannaki, Srikanth V. Krishnamurthy, Michalis Faloutsos, Vivek Mhatre ACM MOBICOM 2007

2. 2 Goal: Improve WLAN network performance Three functions to improve network performance in dense WLANs  Frequency selection  Provides spatial separation between interfering APs  User association  Provides load balancing among APs  Power control  APs shrink their cells to facilitate higher spatial reuse of spectrum

3. 3 Frequency Selection Neighbor Access Points (APs) select different frequencies in order to mitigate interference from each other AP Channel 6Channel 11 Channel 6 Inter-cell Contention: A cannot transmit because B is transmitting A B Cell of A Cell of B

4. 4 User Association Users select less-loaded Access Points (APs) in order to get more throughput AP C A B

5. 5 Power Control APs shrink their overlapping cells in order to reduce interference and improve spatial reuse AP AB

6. 6 Problem statement Which are the functions that should be applied in a specific scenario? In what sequence should they be applied? High level objective: to maximize a fair notion of aggregate network throughput

7. 7 Previous Work Most studies have tried to optimize one of these functions  No studies on the interdependencies between all three functions A. Mishra, V. Shrivastava, D. Agarwal, S. Banerjee. Distributed Channel Management in Uncoordinated Wireless Environments. MOBICOM 2006. N. Ahmed, S. Keshav. SMARTA: A Self-Managing Architecture for Thin Access Points. CoNEXT 2006. K. Sundaresan, K. Papagiannaki. The Need for Cross-Layer Information in Access Point Selection Algorithms. IMC 2006. A. Mishra, V. Brik, S. Banerjee, A. Srinivasan, W. Arbaugh. A Client-Driven Approach for Channel Management in Wireless LANs. INFOCOM 2006. A. Kumar, V. Kumar. Optimal Association of Stations and APs in an IEEE 802.11 WLAN. NCC 2005. T. Korakis, O. Ercetin, S. V. Krishnamurthy, L. Tassiulas, S. Tripathi. Link Quality Based Association Mechanism in IEEE 802.11h Compliant Wireless LANs. RAWNET 2005. B. Leung, K. Kim. Frequency Assignment for IEEE 802.11 Wireless Networks. VTC 2003. Y. Bejerano, S. Han, L. Li. Fairness and Load Balancing in Wireless LANs Using Association Control. MOBICOM 2004.

8. 8 Contribution We perform an extensive experimental study on Testbed-A We quantify the interplay of the three functions  We employ 3 previously proposed algorithms for these functions  We identify the conditions which make the topology conducive to each one of these functions We develop the MDG framework (Measurement Driven Guidelines) We validate the effectiveness of MDG on a different Testbed-B!  Testbed-B is significantly different from Testbed-A  We observe that MDG provides the best strategy in all cases

9. 9 The Structure of This Talk Background on the 3 algorithms Part 1. Experimental Study on Testbed-A - Derivation of conditions Part 2. Building the MDG framework Part 3. Validating MDG on Testbed-B

10. 10 All 3 algorithms based on Gibbs sampling  Fully-saturated downlink traffic Frequency selection algorithm (FS) [Kauffmann et al. ‘07]  Finds the channel allocation with minimum total interference User Association Algorithm (UA) [Kauffmann et al. ‘07]  Finds the state of minimal potential delay of clients  Depends on AP channel access time, AP-client link quality and number of clients per AP Power Control algorithm (PC) [Mhatre et al. ‘07]  Finds the state of minimal potential delay by jointly tuning P TX and Clear Channel Assessment threshold (CCA) Our choice for the algorithms [1] B. Kauffmann et al. “Measurement-Based Self Organization of Interfering 802.11 Wireless Access Networks”. INFOCOM 2007 [2] V. Mhatre, K. Papagiannaki, F. Baccelli. “Interference Mitigation through Power Control in High Density 802.11 WLANs”. INFOCOM 2007

11. 11 Algorithmic requirements The Access Points (APs):  Measure: channel power, load etc  Exchange: information with other APs  Advertise: information to the clients We require minimal functionality from the client  Clients tune their Tx power, CCA threshold  Clients pick Access Point (AP) for association as per the optimization criteria

12. 12 Implementation and Experimental set-up The 3 algorithms are implemented  for both APs and clients  on Intel 2915 prototype driver and firmware Testbed A : U Cambridge, UK  21 APs, 30 client Technical Characteristics  Nodes: Soekris net4826,  Wireless cards: Intel 2915 a/b/g  5-dBi omnidirectional antennae Experiments late at night, to avoid external interference  Both in 802.11a and 802.11g

13. 13 Experiments at a glance We study each function in isolation  To understand the capabilities of each We study all pairwise combinations  To undestand how each affects the other We study the effect of all three of them Methodology  Activate APs and clients in random order  Apply the algorithms  Run throughput measurements to observe gain due to each combination

14. 14 Should we always apply FS? FS is always beneficial  FS outperforms Random Channel Selection (RCS) by 48% in 802.11a and by 65% in 802.11g Frequency Selection (FS) in isolation

15. 15 User Assoc. (UA) in isolation We observe that the performance due to UA is largely dependent on the level of contention  If the contention among APs is high, there is not much for clients to gain  Less contention = more throughput due to UA, when AP load is not balanced Contention is lower in 802.11a than in 802.11g  UA is more favorable in 802.11a than in 802.11g

16. 16 User Assoc. (UA) and Freq. Selection (FS) We apply FS before UA The combination of UA and FS is always beneficial!  The total network throughput becomes higher than the sum of throughputs in the isolated cases!  Much more in 802.11a than in 802.11g

17. 17 How does topology affect the ability of PC to shrink cells?  Five topological cases Cases where PC improves performance  Case a  AP-client link strong (RSSI>-55 dBm)  AP-AP link weaker by at least k dBm (k = 15 to 20)  Case b  Both AP-AP and AP-client links strong (RSSI>-55 dBm) Reduction in power not feasible Increasing CCA makes APs ignore each others’ transmissions --> parallel transmissions possible Power Control helps in some scenarios

18. 18 Power Control does not always help! Cases where PC has no effect  Case c  AP-AP link stronger than AP-client link Isolation is not possible  Case d  AP-client weak and AP-AP even weaker Power reduction reduces the AP-client link quality With CCA increment, AP is disconnected from client  Case e  AP-client link stronger by k dBm than AP-AP link, k < 15 dBm No cell isolation is possible

19. 19 Power control (PC) and Frequency Selection (FS) PC usually does not provide benefits without FS  Many co-channel links under cases (c) and (e)  With FS, remaining co-channel APs have reduced AP-AP link qualities FS + PC is more beneficial in 802.11g than in 802.11a  After FS there is still significant contention in 802.11g, due to fewer channels

20. 20 Power Control (PC) and User Association (UA) PC in conjunction with UA, is usually not beneficial without FS!  UA may create long AP-client links  As long as a user discovers a lightly loaded AP that is going to provide lower delays, the client will associate to that AP  This reduces the AP-client link quality (RSSI) even more AP

21. 21 Applying all 3 functions Blindly applying all three algorithms may hurt the performance ! 24% throughput degradation when applying all 3 algorithms!!

22. 22 The need for a systematic approach We develop MDG (Measurement Driven Guidelines)  A framework for deciding when to apply each function  Based on the empirical observations  Measurement-based inputs:  Whether overlapping cells exist, so as to apply FS  Whether overloaded APs exist, so as to apply UA  Whether AP-AP and AP-client links are conducive for PC Intuitively, MDG:  First mitigates interference  Second balances the load

23. 23 Building MDG Initial steps: 1. Check if FS is beneficial 2. If not, check if UA is beneficial

24. 24 Checking for contention after FS If FS resolves all interference, PC is not needed  If after FS there still exists contention among APs, then further steps depend upon whether the network employs 802.11a or 802.11g.

25. 25 802.11a: UA or PC ? In 802.11a, FS+UA is more beneficial than FS+PC  Applying FS almost eliminates cell overlaps  When contention is limited, it is preferable to apply UA rather than PC

26. 26 802.11g: UA or PC ? In 802.11g, FS+PC is more beneficial than FS+UA  FS does not eliminate cell overlaps, due to the limited number of channels  FS+PC boosts the network performance

27. 27 MDG: The Final Flow Diagram

28. 28 So… The MDG diagram looks really cool, but does it really work? :-)

29. 29 Validating MDG on TestBed-B We validate MDG on a second, different network  UC Riverside Wireless Testbed  Different scale and environmental factors  8 APs, 20 clients Validation procedure  Apply MDG  Evaluate the performance

30. 30 Validating MDG: 802.11a MDG discovers the path that provides the highest total network throughput in 802.11a

31. 31 Validating MDG: 802.11g MDG discovers the path that provides the highest total network throughput in 802.11g

32. 32 Further Testing of MDG Performance Does external interference affect MDG?  MDG performs well even in the presence of other LANs (during daytime) MDG is better than any random network configuration tested  Random channel selection  Random client affiliation  Random P TX and CCA, constant C = P TX * CCA MDG provides the best performance, compared to 40 other random configurations

33. 33 Conclusions MDG maximizes the synergy between frequency selection, user association and power control MDG is measurement driven  Relying on the fundamental understanding of the inter-dependencies between the three functions/algorithms  Grounded on conditions that make the topology conducive to each function We validate the efficiency of MDG on a different network MDG is useful for network management in WLANs in practice

34. 34 Thank you! Questions ?