1. Multiple Access and High Density 802.11 Wireless Access Networks Dina Papagiannaki Intel Research Cambridge

2. November 15th 2006 Konstantina Papagiannaki2 Multiple Access In broadcast environments we need a mechanism to coordinate access among devices (Ethernet, Wireless LANs, Cellular networks) Every transmission is overheard by all other devices in range Simultaneous transmissions lead to collisions that waste network resources Two primary ways of mediating access: Centralized Distributed Design goal: Maximize the number of messages Minimize a station’s waiting time

3. November 15th 2006 Konstantina Papagiannaki3 Centralized vs. Distributed Centralized scheme One node is assuming the role of the master node and determines the order by which slave nodes access the medium. May lead to low medium utilization. Distributed scheme All nodes are equivalent and can talk to each other. Need to coordinate access in order to avoid collisions.

4. November 15th 2006 Konstantina Papagiannaki4 Circuit-mode vs. packet-mode Such a choice depends on the intended workload Circuit-mode allocates part of the medium to a source for its exclusive use – cellular network Packet-mode operates on a per-packet basis, more appropriate for bursty, non-persistent traffic types.

5. November 15th 2006 Konstantina Papagiannaki5 Further constraints Spectrum scarcity Radio channel impairments –Fading – degradation of the signal due to the environment –Multipath interference – reception of signal along multiple paths that may interfere and potentially cancel each other out –Hidden terminal – a transmission may not be overheard by all potentially interfering stations –Capture – the strongest signal at the receiver may be properly decoded (strongest sender has captured the receiver)

6. November 15th 2006 Konstantina Papagiannaki6 Applying those concepts to 802.11 wireless networking IEEE 802.11 is used for Wireless LANs -Carrier Sense Multiple Access with Collision Avoidance (CSMA/CA) - Three variations – 802.11b at 2.4GHz and 11 Mbps, 802.11g at 2.4 GHz and 54 Mbps, 802.11a at 5 GHz and 54 Mbps -Channel impairments dealt using rate adaptation -Different modulation and coding schemes employed that result in different effective transmission rates - Hidden terminal mitigation using Request to Send/Clear to Send (RTS/CTS) control frames

7. November 15th 2006 Konstantina Papagiannaki7 The 802.11 MAC protocol Carrier Sense Multiple Access with Collision Avoidance (CSMA/CA) Before a transmission sender senses the medium If the energy level lower than Clear Channel Assessment (CCA) threshold – medium idle If not, medium busy When the sender wishes to transmit it randomly draws a waiting time [0, CWmin] Each idle slot allows the sender to reduce its CW by 1 slot Upon each unacknowledged transmission the sender doubles its CW up to CWmax (back off)

8. November 15th 2006 Konstantina Papagiannaki8 Spectrum scarcity 802.11 success primarily due to low cost and no licensing fees to use the 2.4 Ghz and 5 GHz bands Small number of operating frequencies in 802.11b/g – slightly more in 802.11a Sharing with non 802.11 devices (microwaves, cordless phones, BT devices, etc.)

9. November 15th 2006 Konstantina Papagiannaki9 The effect of contention In a single contention domain each sender has an equal probability of accessing the medium The greater the number of senders the smaller the throughput Mechanisms for robustness to errors may lead to smaller effective transmission rates Nominal transmission rate of 802.11a/g: 54 Mbps, effective ~30 Mbps, lowest encoding rate 1 Mbps, under contention even lower…

10. November 15th 2006 Konstantina Papagiannaki10 Research Challenges Density of 802.11 APs increases in urban areas Low cost/ease of deployment No coordination in deployment May feature manufacturer default settings Campus and enterprise networks go wireless Higher density could lead to better performance Network management is an ART, especially due to medium dynamics There is no equivalent to over-provisioning Adding APs may be counter-productive

11. November 15th 2006 Konstantina Papagiannaki11 Self-organization in 802.11 networks Tuneable knobs AP frequency (frequency selection) Association of clients to APs (user association) Transmission power and CCA threshold (Power control/MAC layer tuning) Performance evaluation Requirements from existing platforms Further Challenges

12. November 15th 2006 Konstantina Papagiannaki12 The problem statement High density 802.11 wireless networks suffer from sub- optimal performance due to their static configuration (i.e. maximum transmission power, default operating channel, default aggressiveness to access the medium)! AP client

13. November 15th 2006 Konstantina Papagiannaki13 Self-organization objectives Develop fully decentralized algorithms for the self-organization of infrastructure 802.11 wireless networks Seeking mechanisms that aim to optimize global performance using local information alone Robust to changes in the medium and the topology Can be implemented using today’s technology

14. November 15th 2006 Konstantina Papagiannaki14 3 facets to the problem A. Frequency Selection by APs Identify the appropriate frequency to use so as to minimize overall interference across the network B. User association When user association is flexible, balance the users across APs so as to maximize the long-term overall network capacity C. Transmission Power and Aggressiveness to access the medium (CCA threshold) Identify the appropriate level of transmission power and CCA threshold for the APs and clients so as to maximize overall network capacity

15. November 15th 2006 Konstantina Papagiannaki15 Frequency selection formulation P1 P2 Measure noise Measure power

16. November 15th 2006 Konstantina Papagiannaki16 A. Frequency Selection Minimize interference by operating on orthogonal frequencies. Minimize overlap when required to reuse a frequency. Mbps/11gClient1Client2Client3 Before11.816.8614.37 After30.51 (*3)30 (*5)29.45 (*2)

17. November 15th 2006 Konstantina Papagiannaki17 User throughput Internet 1.Channel access time 2.Aggregated transmission delay 3.Wireless channel quality State of the art can lead to unnecessarily low throughput!

18. November 15th 2006 Konstantina Papagiannaki18 Analytical model When wireless the bottleneck… … traffic is downlink – APs are the only senders in the medium … fully saturated traffic conditions – interference caused by APs does not depend on the #clients All clients receive the same long-term throughput if rate adaptation employed In a reference period of time T ET

19. November 15th 2006 Konstantina Papagiannaki19 User association formulation

20. November 15th 2006 Konstantina Papagiannaki20 B. User association Balance the user associations for minimal potential delay fairness. Users take into account the personal and social cost of different association rules. Mbps/11gClient1 Before~ 5 After~ 8

21. November 15th 2006 Konstantina Papagiannaki21 Overall network fairness improved Mean:1428, variance:4378031 Mean:1559, variance: 627638

22. November 15th 2006 Konstantina Papagiannaki22 Implementation on Intel 2915ABG AP Capacity (APC) - MAC Modify firmware to compute fraction of access time, i.e. number of busy slots in a reference period of time (M(a)) Nominal capacity given by 11a/b/g (C(a)) Aggregated transmission delay (ATD) – MAC/PHY Modify firmware/ucode to measure amount of time between queueing the packet towards a client and the reception of the ACK (rate scaling, and retransmissions) Keep a list of client MACs and delay, compute sum of delays Transmission rate for new client approximated using RSSI - PHY APC/ATD advertized through Beacon frames

23. November 15th 2006 Konstantina Papagiannaki23 Experimental Results AP1AP2AP3 Ch10 Ch3 C1C3C2 4 Mbps

24. November 15th 2006 Konstantina Papagiannaki24 Power Control in 802.11 Heterogeneous transmit powers across nodes can lead to node starvation! 1 st order starvation We need to ensure that there is symmetry in the nodes’ contention domains.

25. November 15th 2006 Konstantina Papagiannaki25 What is the benefit of power control? Reducing transmission power can reduce interference in the network Increasing transmission power can improve client SINR thus allowing for higher transmission rates There is a tradeoff between the amount of interference we introduce in the network and the additional throughput benefit at the client

26. November 15th 2006 Konstantina Papagiannaki26 Condition for starvation free power control We need to ensure network symmetry We have proven that for starvation-free power control we need to keep the product of CCA threshold and transmission power constant CCA * P = C The louder you are going to shout the more carefully you should listen for the nodes that whisper

27. November 15th 2006 Konstantina Papagiannaki27 How do we maximize network capacity? We need to identify these values of C that result in the greatest transmission concurrency We can optimize C using Gibbs sampling in order to maximize network capacity Input: channel gains between APs, channel gains from AP to clients, number of clients per AP, transmission power Output: APs select transmission power and CCA. Clients follow the setting of their AP.

28. November 15th 2006 Konstantina Papagiannaki28 Experimental Testbed

29. November 15th 2006 Konstantina Papagiannaki29 C. Power Control / CCA adaptation Tune power to offer the best transmission rate to the farthest client while not introducing excessive interference to neighboring co-channel devices. Adjust CCA to increase transmission concurrency across the network. Mbps/11gClient1Client2Client3 Before11.816.8614.37 After29.45 (*3)22.59 (*4)30.51 (*2)

30. November 15th 2006 Konstantina Papagiannaki30 Experimental Results GainDefaultCCA Client SS03: 149%15% Client SS15:228%34% Client SS24: 112%3%

31. November 15th 2006 Konstantina Papagiannaki31 Simulation Results (topology – 8 APs, 26 STAs, 802.11a, AP-STA: 3.5m)

32. November 15th 2006 Konstantina Papagiannaki32 Simulation Results (power, CCA)

33. November 15th 2006 Konstantina Papagiannaki33 Simulation Results (throughput)

34. November 15th 2006 Konstantina Papagiannaki34 Summary Results Gibbs also leads to the use of a smaller transmission power that can extend client’s lifetime

35. November 15th 2006 Konstantina Papagiannaki35 Implementation Requirements AP Capacity (MAC) Aggregated Transmission Delay (PHY/MAC) Number of users Worst Client Channel Gain (PHY) Introduction of new Beacon fields (CCA, TxPower, auxiliary variables) Channel Switch Announcements (802.11h/DFS/TPC) Measurements (802.11k/802.11e)

36. November 15th 2006 Konstantina Papagiannaki36 Larger Scale Experimentation SWAN testbed at William Gates Building 80 Soekris dual mini-PCI boards Intel 2915 ABG cards with modified ucode/firmware PoE switches for ease of manageability Investigation of benefits of the three different algorithms compared to the state of the art and their incremental benefits

37. November 15th 2006 Konstantina Papagiannaki37 Floorplans Ground floor 1 st floor 2 nd floor

38. November 15th 2006 Konstantina Papagiannaki38 More exciting problems….

39. November 15th 2006 Konstantina Papagiannaki39 Community mesh networking Internet

40. November 15th 2006 Konstantina Papagiannaki40 Questions?

41. November 15th 2006 Konstantina Papagiannaki41 BACKUP