380 likes | 494 Views
Throughput Improvement in 802.11 WLANs using Collision Probability Estimates. Avideh Zakhor E. Haghani , M. Krishnan, M. Christine, S. Ng Department of Electrical Engineering and Computer Sciences U.C. Berkeley October 2010. Outline. Background Type of loss in wireless networks
E N D
Throughput Improvement in 802.11 WLANs using Collision Probability Estimates AvidehZakhor E. Haghani, M. Krishnan, M. Christine, S. Ng Department of Electrical Engineering and Computer Sciences U.C. Berkeley October 2010
Outline • Background • Type of loss in wireless networks • Estimating collision probabilities two years ago • Using estimates to improve throughput • Modulation rate adaptation last year • This year: • Carrier sense threshold • Packet length adaptation • Experimental verification
Motivation & Goal • Improve throughput: • Differentiate between various loss events • Estimate probability of occurrence of each type • Adapt: • Link adaptation algorithm • Packet length • Carrier sense threshold • Contention window • Transmit power • FEC
Types of Loss 802.11 Network DCF – contention window Direct Collision (DC): nodes start transmitting in same slot Hidden Terminal Staggered Collision: one node starts transmitting in the middle of another node’s packet SC1: node in question is first SC2: node in question is second Channel Errors Large pathloss due to distance/obstacles (large timescale) Random multipath fading (small timescale) B A AP 4
Estimating Collision Probability • Each node/AP collects binary-valued ‘busy-idle’ (BI) signal • 1 when local channel is occupied, 0 otherwise • AP broadcasts its BI signal periodically ~14kb/s, 3% overhead • Nodes use their BI signal along with AP’s to estimate PC C A AP1 AP2 B Node A: AP1: Node B: Krishnan, Pollin, and Zakhor, “Local Estimation of Probabilities of Direct and Staggered Collisions in 802.11 WLANs”, IEEE Globecom 2009.
What to do with these estimates? Link adaptation: Current techniques assume all losses are due to channel error lower rate unnecessarily Make staggered collision problem worse longer packets Adaptive packetization: if most collisions are staggered due to hidden nodes, need shorter packets Joint throughput optimization of: Modulation rate Packet length FEC Contention window Retransmit limit Transmit power Carrier sensing threshold Use of RTS/CTS Optimization might be different for delay Fairness issues 6
Outline • Background • Type of loss in wireless networks • Estimating collision probabilities two years ago • Using estimates to improve throughput • Modulation rate adaptation last year • This year: • Carrier sense threshold • Packet length adaptation • Experimental verification
Carrier Sense Optimization in 802.11 • CSMA network - nodes transmit only if sensed power < CS threshold • Trade-off between hidden node problem and exposed node problem • CS threshold => # of hidden nodes , # of exposed nodes • Tune CS threshold to: • minimize # of hidden nodes + # of exposed nodes for the transmitter • Increase throughput • A (the Station) is transmitting to B (the AP). • : transmission range -- Signal can be decoded • : CS range -- Received power > CS threshold) • : interference range -- Any transmission in this range collides with A’s signal at B • E is an exposed node and F is a hidden node to A.
Busy/Idle Signal • AP broadcasts its BI signal, BIAP, every Δ seconds • Each station records multi-leveled sensed energy level for the same period of Δ seconds • Station generates its own BI signal • Depends on CS threshold ϒ. • For p, q ∈ {0, 1},
Hidden and Exposed nodes in BI signal • Hidden node problem: BISTA = 0 and BIAP = 1 => collisions • Exposed node problem: BISTA = 1 and BIAP = 0 => excess backoff • Continuous-valued sensed power depends on other nodes sending, but node can affect binary-valued BISTA by adapting CST • BISTA = 1{power > CST} • Adapt to minimize + , or Exposed node transmission Hidden node transmission
Optimization Function • Hidden and exposed nodes reduce the throughput • Can affect number of hidden and exposed nodes by tuning ϒ |Transmissions of Hidden Nodes| ∝ |Transmissions of Exposed Nodes| ∝ • Optimization: where • As increases: • P10 decreases – fewer exposed nodes • P01 increases – more hidden nodes
Algorithm • Record energy level of the channel for Δ=3 seconds. • Receive BI signal from AP. • Calculate the value of function F for all possible values of carrier sense threshold. • Find the value of the carrier sense threshold that minimizes F. • Find the value of F for the previous value of carrier sense threshold. • If the difference is more than 5% of previous value change the carrier sense threshold.
Simulation Setup • 7 APs, 50 nodes • APs have fixed CST for each simulation • Different over various simulations • 2 methods for comparison: • Nodes have same fixed CST as APs • Nodes asynchronously adapt using our algorithm: • Use current CST for 3+ seconds, where is random • Solve optimization for data from most recent 3 seconds • Consider all nodes in 10 different 60-second simulations with different topologies 500 total nodes • Repeat this for each value of AP CST
Simulations: Aggregate Throughput vs AP CST • Up to 50% total throughput improvement • Moderate decrease when AP CST is very low – single collision domain • The average of log-throughput is increased in all scenarios => adaptive CST algorithm behaves fairly.
Simulation Results: Node Throughput • 80% of nodes gain throughput, only 10% lose • Median: 81%, Mean: 131% • Improvement depends on locations of hidden and exposed nodes
Simulation Result: Attempts and Losses • Adaptive algorithm results in: • Lower loss probability • Fewer transmission attempts • More efficient channel use
Outline • Background • Type of loss in wireless networks • Estimating collision probabilities two years ago • Using estimates to improve throughput • Modulation rate adaptation last year • This year: • Carrier sense threshold • Packet length adaptation • Experimental verification
Effects of MAC Layer Packet Length Impact of packet size on effective throughput • Protocol header overhead • Larger packet size is preferable • Channel fading • Smaller packets are less vulnerable to fading errors • Direct collisions • Direct collision probability is independent of packet size • Staggered collisions in presence of hidden terminals • Smaller packets are less susceptible to collide with transmission from hidden terminals
Packet Loss Model • Pure BER-based • Used in length adaptation literature • Assume constant BER over all packets over all time • Simple analysis • Does not account for packet-to-packet channel variation • BER-SNR • Assume constant BER over each packet • Assume distribution on SNR: Rayleigh, Log-Normal, Rice • BER known function of SNR and modulation rate • Accounts for channel variation • Pure BER is special case where SNR distribution is delta L = payload length Lh = header length Rp = payload modulation rate Rh = header modulation rate f () = distribution of SNR BER() functions are known
Single-Node Throughput vs Length as a function of BER-SNR Variance Optimal packet length increases with SNR variance
Approach: Gradient Search TP = throughput L = packet length sendFreq =# packets/sec PSC1 = P(SC1) Pe = P(channel error) C’ constant • Gradient of TP w.r.t. packet length: • Pe estimated as: L known; sendFreq and PL empirical counting, m2 and Pc [1] next page
Estimating where = P(error for packet with SNR ) = P(header error for packet with SNR ) • Estimate Pe from [1] look up • Assume single parameter or two parameter distribution
Algorithm • Observe for N seconds without adaptation, • Estimate • Adjust L by where is adjusted as follows:
Verification of via NS Simulations • Scenario: 7 Aps, 50 nodes, all using constant packet length • Vary L for a single node to examine TP vs L • Locally compute and compare to slope of empirical TP vs L curve Node 1 Node 2
Example of Adapted Length and Throughput Change • 7 APs, 50 nodes, -89 dBm noise • Periphery nodes choose shorter lengths • Spatial correlation between gain/loss • Highest % gain in T.P lowest absolute T.P. nodes Length % throughput change Total throughput =gain =loss =standard =adaptive
Throughput Improvement vs Noise Power • High noise power High Pe more nodes choose smaller L -89 dBm -95 dBm
Outline • Background • Type of loss in wireless networks • Estimating collision probabilities two years ago • Using estimates to improve throughput • Modulation rate adaptation last year • This year: • Packet length adaptation • Carrier sense threshold • Experimental verification
Experimental Verification of Pc Estimation • Implemented mechanism behind collision probability estimation technique using Ath5k open source wireless card driver • Topology: • Node 1 sends to AP 1, and computes estimates • Node 2 sends to AP 2 to cause hidden node collisions • Sniffers observe ground truth
Estimation Approach – ‘Busy-Idle’ Signal • Each node/AP collects binary-valued ‘busy-idle’ (BI) signal • 1 when local channel is occupied, 0 otherwise • Also collect TX signal - 1 when transmitting, 0 otherwise C A AP1 AP2 B Node A: AP1: Node B:
System Design • 4 steps: • Collect available carrier sense data from wireless card • Process this data to generate BI and TX signals • Align BI and TX signals of station and AP • Compute estimates • Ideally completely implemented at driver level • Current implementation only collects data in real time • Data is processed offline in MATLAB
Collecting Carrier Sense Data • Access “profile count” registers; observe this behavior: • AR5K_POFCNT_CYCLE: constantly incrementing like clock • AR5K_PROFCNT_TX: increasing at same rate at CYCLE when transmitting, constant otherwise • AR5K_PROFCNT_RXCLR: increasing at same rate at CYCLE when channel is occupied, constant otherwise • In theory: • BI signal is slope of RXCLR vs CYCLE • TX signal is slope of TX vs CYCLE • Practically: can capture • sequentially – not simultaneously • not necessarily regularly • “time” of TX or RXCLR sample is bounded by value of previous and subsequent sampled value of CYCLE
Generating BI/TX signals Candidate busy section is set of consecutive y-values (RXCLR or TX) which are strictly increasing: Equation 1: +lower bound ×upper bound b
Aligning Station and AP signals • To estimate collision probability, need to line up TX and BI signals between station and AP • Scale to adjust for different clock speeds • Use large scale view of packet start times • Align TX signals more sparse than BI signals; easier • AP TX consists of ACKS, some of them to station • Line up inter-packet times • BI signals follow since they are collected on same clock as TX • Most packets aligned within 40s of each other
Experimental Setup • Topology: • Node 1 sends to AP 1, and computes estimates • Node 2 sends to AP 2 to cause hidden node collisions • Sniffers observe ground truth • Variables: • Transmit power of node 1 • to affect Pe • Sending rate of node 2 • to affect Pc
Experimental Results • 75 total estimates: • 5 levels of Pc with 15 estimates each: • 6 estimates with Pe~0 • 9 estimates with 20<Pe<40 =low Pe X =high Pe Pc estimates are within 5% accuracy
Future Work: Contention Window Adaptation • Contention Window Adaptation strategy: • Nodes wait for random number, drawn uniformly from {1,2,…,W} of idle time slots before transmitting • If packet fails, WW • By default =2 • Can show this is asymptotically optimal as n for single collision domain with no fading/noise, i.e. all losses are DCs • What happens when we include other types of losses in the model? • E.g. if all losses are due to channel, want =0 • What about more general schemes where we can choose arbitrary distributions for backoff time?
Future Work: Delay-sensitive traffic • Effective throughput – throughput received within delay bound • What bit rate & retransmit limit (γ) & delay limit (τ) maximize the effective throughput (η)? • Derive an analytical expression/model for effective throughput • Use the BI signal information • Nodes make observations to estimate parameters of model • Advantages: • Can adapt fast in multi-dimensional parameter space • Preferable to making one parameter at a time observations of throughput
Future Work: Application to asymmetric TCP • Links 1,4 subject to network congestion • Link 2 subject to channel errors • Link 3 subject to channel errors AND collisions • TCP assumes symmetric channel only limited by congestion • Question: Can we take advantage of knowing collision probability to adjust parameters of asymmetric TCP algorithms? • low Pc => channel is roughly symmetric • higher Pc => increased asymmetry? Internet 1 2 4 3 AP client server asymmetry