Traffic-Aware Channel Assignment in Enterprise Wireless LANs by kpr16177

VIEWS: 18 PAGES: 27

									  Traffic-Aware Channel
 Assignment in Enterprise
      Wireless LANs
  Eric Rozner University of Texas at Austin
 Yogita Mehta University of Texas at Austin
Aditya Akella University of Wisconsin-Madison
     Lili Qiu University of Texas at Austin
              IEEE ICNP 2007
               October 18, 2007
                   Motivation
• Increasing campus & enterprise WLAN
  popularity
  – Laptops, smart phones, wireless gaming consoles,
    etc
• Increased density and usage → interference
• Limited number of non-overlapping channels
  – 802.11b and 802.11g only have 3 (1, 6, and 11)
  – Not always feasible to assign non-overlapping
    channels

                                                     2
                          Related Work
 • Previous channel assignment schemes
    –   Manual configuration [Grier]
    –   Maximize RSS at expected high-demand points [Lee02]
    –   Client-side interference [Mishra06]
    –   Commercial products [AutoCell, AirView]
         • No public information due to proprietary nature

Approaches assume network traffic is static or uniform!

 • Wireline traffic engineering
    – Benefits of traffic-awareness [Awduche99, Awduche02,
      Xiao00]
    Our Contribution: Effective channel assignment
schemes that adapt to prevailing WLAN traffic demands 3
               Motivating Example
 Traffic-Agnostic                     Traffic-Aware
 Throughput: 10 Mbps            Throughput: 15 Mbps
Demand(a) = 5 Mbps              Demand(b) = 5 Mbps
 Channel 1             a      b          Channel 6
              2.5 Mbps
 Throughput: 5 Mbps              Throughput: 5 Mbps


          Traffic-aware channel assignment
                   can be beneficial!
                     c          d
Demand(c) = 0 Mbps               Demand(d) = 5 Mbps
         11
 Channel 6                                Channel 1
                                         Channel 11
 Throughput: 0 Mbps               Throughput: 5 Mbps
                                 Throughput: 2.5 Mbps

                                                        4
Traffic-Aware Framework
     Measure interference graph
       Obtain traffic demands
       from previous interval
 Predict demands for current interval

       Compute traffic-aware
        channel assignment

        New assignment≠                 No
         old assignment
                    Yes
    Change channel assignment
                                             5
    Key Questions to Achieve Traffic-
      Aware Channel Assignments
• How to develop traffic-aware channel
  assignment algorithms?
• How to estimate traffic that varies over time?
• How to estimate the interference graph?
• How to handle non-binary interference?
• How to efficiently change channels?
• How much does traffic-awareness improve
  network performance and when is it
  beneficial?                                      6
               Traffic-Awareness
• Weigh interference metric by traffic demands
  – SA - Node A’s sending demands
  – RA - Node A’s receiving demands
• W A,B = SA×SB + SA×RB + SB×RA
  – 1st term: sender-side interference
     • 802.11 MAC is CSMA/CA: One sender at a time
  – 2nd and 3rd terms: interference at receivers
     • Collisions increase loss, contention window



                                                     7
           Channel Separation Metric
 • SepA,B = min(|chan(A) - chan(B)|, 5) if A, B interfere
         = 5 otherwise

 Metric      Traffic-agnostic               Traffic-aware

 Client-            Max:                          Max:
agnostic       ∑i,j ∈AP Sepi,j           ∑i,j ∈AP W ij × Sepi,j
                    Max:                          Max:
Client-     ∑i,j ∈AP∪Clients Sepi,j   ∑i,j ∈AP ∪Clients W ij×Sepi,j
aware

 • Traffic-awareness can be applied to other metrics
 • Finding optimal solution is NP-Hard [Mishra06]                 8
    Obtaining Channel Assignments
• Initialization algorithm
  – Inspired by Chaitin’s approach to register
    allocation problem [Chaitin82]
  – Basic notion: Wait to assign channels of APs with
    many conflicts b/c such assignments are more
    important
• Simulated annealing to improve initial
  assignment
  – Randomly change channel of one AP and its
    clients
  – If metric improves, select current assignment;
      If not, select it with some non-zero probability P   9
  – Probability P decreases as # iterations increases
    Key Questions to Achieve Traffic-
      Aware Channel Assignments
• How to develop traffic-aware channel
  assignment algorithms?
• How to estimate traffic that varies over time?
• How to estimate the interference graph?
• How to handle non-binary interference?
• How to efficiently change channels?
• How much does traffic-awareness improve
  network performance and when is it
  beneficial?                                      10
       Estimating Traffic Demands
• Measure past traffic demands
  – Most commercial APs export SNMP interface
  – SNMP provides demands in 5 min intervals
• Predict current demands based on history
  – EWMA: Exponentially-weighted moving average
  – PREV: Use previous interval’s demands
  – PREV_N: Find channel assignment that’s
    optimized over past N intervals
  – PEAK_N: Find channel assignment that’s
    optimized over the worst case in past N intervals.
                                                         11
    Key Questions to Achieve Traffic-
      Aware Channel Assignments
• How to develop traffic-aware channel
  assignment algorithms?
• How to estimate traffic that varies over time?
• How to estimate the interference graph?
• How to handle non-binary interference?
• How to efficiently change channels?
• How much does traffic-awareness improve
  network performance and when is it
  beneficial?                                      12
   Estimating the Interference Graph
• Measure max throughput on any 2 links [Padhye05]
  –   A’s max broadcast rate when it sends alone
  –   A’s max broadcast rate when it sends with node B
  –   BR = Total throughput together/Total throughput alone
  –   BR close to 0.5 → A, B interfere (take turns sending),
         close to 1.0 → A, B don’t interfere
• Estimate max throughput on any 2 links via an
  interference model [Reis06]
• Estimate max throughput on any set of links via a
  general interference model [Qiu07]
• Use coordinated probing [Ahmed06]
• Further improvement of interference graph
  estimation directly benefits our channel assignment 13
    Key Questions to Achieve Traffic-
      Aware Channel Assignments
• How to develop traffic-aware channel
  assignment algorithms?
• How to estimate traffic that varies over time?
• How to estimate the interference graph?
• How to handle non-binary interference?
• How to efficiently change channels?
• How much does traffic-awareness improve
  network performance and when is it
  beneficial?                                      14
         Non-Binary Interference
• Interference can be non-binary in practice
  – Variations in RSS cause intermittent interference
  – SNR under one sender ≥ SNR_Threshold
  – SNR under two (or more) senders ≤
    SNR_Threshold
• Extend the channel assignment metric to
  handle non-binary interference
  – Degree of interference is weighed by the
    throughput reduction based on BR

                                                        15
    Key Questions to Achieve Traffic-
      Aware Channel Assignments
• How to develop traffic-aware channel
  assignment algorithms?
• How to estimate traffic that varies over time?
• How to estimate the interference graph?
• How to handle non-binary interference?
• How to efficiently change channels?
• How much does traffic-awareness improve
  network performance and when is it
  beneficial?                                      16
             Channel Switching
• Switching delay - hardware (AP & client)
  – 200μs Intel ProWireless
  – 10-20ms Netgear Atheros, Cisco Aironet,
    Prism 2.5
• Re-association delay - software (client
  only)
  – Default: clients scan all channels to assoc.
    • Scanning time dominates (100’s of ms
      [Ramani05])
  – Explicit Notification: APs broadcast channel
    • Can send multiple times to protect against loss17
    Key Questions to Achieve Traffic-
      Aware Channel Assignments
• How to develop traffic-aware channel
  assignment algorithms?
• How to estimate traffic that varies over time?
• How to estimate the interference graph?
• How to handle non-binary interference?
• How to efficiently change channels?
• How much does traffic-awareness improve
  network performance and when is it
  beneficial?                                      18
          Evaluation Methodology
• NS-2 Simulation
  – Synthetic traces: when traffic-awareness is
    beneficial
  – Trace-driven simulations: more realistic settings
    • SNMP data from Dartmouth 2004 and IBM 2002 traces
  – 1024 UDP packet + fixed rate
• Testbed Experiments
  – 25 nodes (MadWifi, 802.11g); 2 floors of office
    building
    • Run at night to avoid interference from resident WLAN
  – Empirically measure non-binary interference graph
  – Study TCP/UDP and fixed rate/auto rate          19
                  Synthetic Results
• Uniform: AP demands uniform over [0:MAX]
• Hotspot: Pick 1 AP & all other APs in range as a
  hotspot,
  Hotspot APs uniform: [0:MAX]; others: [0:LOW]

                                        20% of runs:
                                    At least 33% improv
         20% of runs:
     At least 8.5% improv




Higher benefit when traffic-distribution is more uneven20
               Trace-Driven Results
• Compare against client-agnostic/traffic-agnostic
  baseline
• Average improvements against baseline over 3
  buildings:
   – Traffic-aware, client-agnostic: 5.2-11.5%
   – Traffic-aware, client-aware: 8.3-12.8%




Traffic-awareness provides benefits under real demands
                                                     21
             Prediction Results
   M.A.E.      EWMA       PREV         PEAK2
   ResBldg     0.48       0.49         0.70
   LibBldg     0.43       0.47         0.57

Prediction error can be high due to low aggregation




      Prediction algorithms still perform well
                                                      22
            (EWMA usually within 6%)
               Testbed Results
• TCP results shown, error bars denote standard
  deviation
• Zipf-like slope (X-axis) generates demands
  – Higher slope → more uneven the demands




          Traffic-awareness beneficial for
           both fixed-rate and multi-rate         23
       Channel Switching Overhead
• Measure AP-Client throughput over a 10 minute
  transfer
  – Vary frequency of switching AP’s channel
  – Examine different levels of client activity




 Overhead is minimal for ≥ 2 min switching interval   24
                           Conclusion
• Main contributions
  – Traffic-aware channel assignment algorithms in WLANs
  – Considered several practical issues
     •   Measure wireless interference
     •   Cope with realistic wireless interference patterns
     •   Measure & predict traffic demands
     •   Minimize the overhead of channel switching
  – Extensive evaluation via simulations and experiments
     • Traffic-awareness benefits under uneven demand distribution
     • Traffic-awareness benefits TCP/UDP and Fixed/Multi-Rate
• Future work
  – Develop traffic-aware techniques for other wireless
    network operations (e.g. power control, routing)
                                                                     25
                Questions?

•Thanks!
 – Eric Rozner
 – erozner@cs.utexas.edu




                             26
             Non-Binary Interference
• BR metric review:
  – BR = Total throughput together/Total throughput alone
  – BR close to 0.5 → A, B interfere (take turns sending),
      close to 1.0 → A, B don’t interfere
• Extend the BR metric:
  –   BR = min(1, max(0.5, BR)); //BR in range 0.5 .. 1
  –   LocInterf = 2 − 2 × BR; //map BR to range 0 .. 1
  –   ChannelDiff = min(|Ci − Cj|, 5);
  –   ChannelInterf = 1 − ChannelDiff × 0.2;
  –   OverallInterf = ChannelInterf × LocInterf ;
• Traffic-aware, client-agnostic metric becomes:
  – Min: ∑i,j∈AP W × OverallInterf(i, j) //others follow     27

								
To top