Download presentation

Presentation is loading. Please wait.

Published byByron Hazelett Modified over 2 years ago

1
Control-Theoretic Adaptive Mechanisms for Performance Optimization of IEEE WLANs: Design, Implementation and Experimental Evaluation Author: Paul Horaţiu Pătraş Advisor: Dr. Albert Banchs Roca March 18, 2011

2
18/03/2011 Outline Motivation & Background Centralized Adaptive Control (CAC) Algorithm Distributed Adaptive Control (DAC) Algorithm Experimental Evaluation Conclusions & Future Work Paul Patras 2

3
18/03/2011 IEEE standard employs a CSMA/CA channel access scheme whose performance depends on the Contention Window (CW) Motivation Paul Patras 3

4
18/03/2011 IEEE standard employs a CSMA/CA channel access scheme whose performance depends on the Contention Window (CW) Motivation Paul Patras 4 Underutilized channel

5
18/03/2011 IEEE standard employs a CSMA/CA channel access scheme whose performance depends on the Contention Window (CW) Motivation Paul Patras 5 High collision rate

6
18/03/2011 IEEE standard employs a CSMA/CA channel access scheme whose performance depends on the Contention Window (CW) [1] G. Bianchi, Performance Analysis of the IEEE Distributed Coordination Function, IEEE Journal on Selected Areas in Communications, vol. 18, no. 3, pp. 535–547, March [2] P. Serrano, A. Banchs, P. Patras, and A. Azcorra, Optimal Configuration of e EDCA for Real-Time and Data Traffic, IEEE Transactions on Vehicular Technology, vol. 59, pp. 2511–2528, June Motivation Paul Patras 6 Given the number of active stations and their sending rate, there exists an optimal CW that maximizes performance [1,2]

7
18/03/2011 Motivation IEEE Standard Stations are capable of updating their CW the AP relies on beacon frames to broadcasts the MAC configuration Only a fixed recommended setting is specified suboptimal performance in most scenarios Paul Patras 7

8
18/03/2011 Background Previous works dynamically adjust the CW based on observed network conditions to improve performance Centralized approaches (e.g., [3-5]) – AP periodically computes the EDCA configuration and distributes it to the active stations Distributed approaches (e.g., [6-8]) – stations compute their configuration independently suitable also for ad-hoc mode [3] A. Nafaa and A. Ksentini and A. Ahmed Mehaoua and B. Ishibashi and Y. Iraqi and R. Boutaba, Sliding Contention Window (SCW): Towards Backoff Range-Based Service Differentiation over IEEE Wireless LAN Networks, IEEE Network, vol. 19, pp. 45–51, July [4] J. Freitag and N. L. S. da Fonseca and J. F. de Rezende, Tuning of e Network Parameters, IEEE Communications Letters, vol. 10, pp. 611–613, August [5] Y. Xiao, H. Li, and S. Choi, Protection and guarantee for voice and video traffic in IEEE e wireless LANs, in Proc. IEEE INFOCOM, vol. 3, pp. 2152–2162, March [6] G. Bianchi, L. L. Fratta, and M. Oliveri, Performance evaluation and enhancement of the CSMA/CA MAC protocol for wireless LANs, in Proceedings of PIMRC 96, Taipei, Taiwan, October [7] M. Heusse, F. Rousseau, R. Guillier, and A. Duda, Idle Sense: an optimal access method for high throughput and fairness in rate diverse wireless LANs, in Proceedings of SIGCOMM. New York, NY, USA, August [8] F. Cali, M. Conti, and E. Gregori, IEEE protocol: design and performance evaluation of an adaptive backoff mechanism, IEEE Journal on Selected Areas in Communications, vol. 18, no. 9, September Paul Patras 8

9
18/03/2011 Contributions Limitations of prior works: –Based on heuristics Lack analytical support –Require modifications of the hardware and/or firmware –Their performance has not been assessed with real deployments Advantages of the proposed adaptive algorithms: –Based on analytical models of the WLAN performance –Sustained by single-/multi-variable control theory foundations –Can be implemented by current devices without introducing any modifications into their hardware and/or firmware –Validated under a wide set of conditions outperform previous adaptive schemes / insights on suitability for practical deployments Paul Patras 9

10
18/03/2011 Outline Motivation & Background Centralized Adaptive Control (CAC) Algorithm Data Traffic Scenario Real-Time Traffic Scenario Distributed Adaptive Control (DAC) Algorithm Experimental Evaluation Conclusions & Future Work Paul Patras 10

11
18/03/2011 Centralized Adaptive Control (CAC) Based on a single node (AP), that periodically computes the CW configuration to be used by stations Objective: dynamically adjusts the CW min parameter to maximize the total throughput of data stations Does not require any modifications of the stations compatible with the IEEE standard Does not require estimating the number of active stations and their level of activity (only observes successfully received frames at the AP) Paul Patras 11

12
18/03/2011 Throughput Analysis and Optimization We first consider the case when all the stations are saturated and later relax this assumption According to Bianchi [1], we define the probability τ that a station transmits in a randomly chosen time slot (1) where p is the conditional collision probability: (2) We express the throughput of the WLAN as a function of τ (3) Paul Patras 12

13
18/03/2011 Throughput Analysis and Optimization The optimal value of τ that maximizes throughput (4) The corresponding optimal collision probability (5) Paul Patras 13

14
18/03/2011 Throughput Analysis and Optimization The optimal value of τ that maximizes throughput (4) The corresponding optimal collision probability (5) Under optimal operation, the collision probability is a constant independent of the number of stations Paul Patras 14

15
18/03/2011 CAC Algorithm Paul Patras 15 Reference signal

16
18/03/2011 CAC Algorithm Paul Patras 16 Reference signal If p obs is larger/smaller than the optimal value, trigger an increase/decrease of the CW min, but avoid oscillations

17
18/03/2011 CAC Algorithm Paul Patras 17 Reference signal Increase/decrease in small steps the CW min, but do not act too conservatively

18
18/03/2011 CAC Algorithm We model the WLAN from a control-theoretic standpoint: –The controller the adaptive algorithm running at the AP –The controlled system the WLAN itself Paul Patras 18

19
18/03/2011 CAC Algorithm 1.AP measures the collision probability of the WLAN resulting from the current CW configuration during a beacon interval Paul Patras 19

20
18/03/2011 CAC Algorithm 1.AP measures the collision probability of the WLAN resulting from the current CW configuration during a beacon interval 2.AP computes the new CW min based on the measured collision probability and distributes it to the stations in a new beacon Paul Patras 20

21
18/03/2011 The Controller Well established scheme from discrete-time control theory: Proportional Integrator (PI) controller (6) Takes as input an error signal The AP computes and distributes the CW configuration to the stations (7) CAC Algorithm Paul Patras 21

22
18/03/2011 CAC Algorithm p obs – estimation relies on inspecting the retry flag of the correctly received frames (no changes at the AP) (8) Paul Patras 22

23
18/03/2011 CAC Algorithm We need to characterize WLAN with a transfer function that takes CW min as input and p obs as output nonlinear relationship We linearize it considering the perturbations about the stable point of operation (9) We obtain the following expression (10) Paul Patras 23

24
18/03/2011 Controller Configuration To compute the {K p,K i } configuration of the controller we apply the Ziegler-Nichols rules [9] Guarantee a proper tradeoff between stability and speed of reaction to changes (11) The system is stable with the proposed configuration (Theorem 1) [9] G. F. Franklin, J. D. Powell, and M. L. Workman, Digital Control of Dynamic Systems. Prentice Hall, 3rd ed., Paul Patras 24

25
18/03/2011 Throughput performance CAC follows the static optimal configuration [2] and outperforms the default EDCA configuration Performance Evaluation Paul Patras 25

26
18/03/2011 Validation of the designed controller stable configuration CW min evolves stably with minor oscillations about the optimal point Performance Evaluation Paul Patras 26

27
18/03/2011 Validation of the designed controller stable configuration fast response to changes CW min evolves stably with minor oscillations about the optimal point The system reacts fast to changes with the proposed configuration Performance Evaluation Paul Patras stations 30 stations

28
18/03/2011 Validation of the designed controller unstable configuration slow response to changes A large {K p,K i } setting yields unstable behavior A smaller {K p,K i } setting gains stability but induces slow response Performance Evaluation Paul Patras stations 30 stations

29
18/03/2011 Comparison against other approaches Our solution outperforms other centralized approaches [3,4] based on heuristics, since it guarantees optimal performance in all scenarios Performance Evaluation Paul Patras 29

30
18/03/2011 Performance Evaluation Additional simulation experiments: –Instantaneous throughput upon changes in the network –Coexistence with non-saturated stations –Performance under bursty traffic –Impact of channel errors further validate the performance of the proposed CAC Paul Patras 30

31
18/03/2011 CAC Summary Analyzed the WLAN saturation throughput and derived the collision probability that yields optimal performance Proposed Centralized Adaptive Control algorithm [9,10] –Dynamically adjusts the CW configuration of IEEE stations with the goal of maximizing the overall throughput –Fully compatible with the IEEE standard –Based on a PI controller achieves a good tradeoff between stability and speed of reaction to changes [9 ] P. Patras, A. Banchs, and P. Serrano, A Control Theoretic Approach for Throughput Optimization in IEEE e EDCA WLANs, Mobile Networks and Applications (MONET), vol. 14, pp. 697–708, December [10] P. Patras, A. Banchs, and P. Serrano, A Control Theoretic Framework for Performance Optimization of IEEE Networks, in IEEE INFOCOM Student Workshop, (Rio de Janeiro, Brazil), pp. 1–2, April Paul Patras 31

32
18/03/2011 Outline Motivation & Background Centralized Adaptive Control (CAC) Algorithm Data Traffic Scenario Real-Time Traffic Scenario Distributed Adaptive Control (DAC) Algorithm Experimental Evaluation Conclusions & Future Work Paul Patras 32

33
18/03/2011 Real-Time Traffic Scenario EDCA parameters for video traffic lead to suboptimal behavior when the WLAN is heavily loaded CAC maximizes the total throughput but does not consider the delay requirements of real-time applications We extend CAC with the goal of optimizing the WLAN performance under video traffic The proposed CAC-VI minimizes average delay and improves QoE Paul Patras 33

34
18/03/2011 Analytical Model We analyze the delay of a WLAN under video traffic and compute the collision probability that provides optimal performance Set CW min dynamically to drive the collision probability in the WLAN to the optimal value that minimizes the average delay Use the maximum value allowed for the TXOP parameter Do not perform BEB upon a collision ( CW min = CW max ) Paul Patras 34

35
18/03/2011 Analytical Model We analyze the WLAN based on a Markov chain where at state i there are i backlogged stations Key assumptions of the analysis: –Aggregate arrivals follow a Poisson process ( λ = n· λ i ) –Access delays are exponentially distributed –Stations do not accumulate more than one video frame Paul Patras 35

36
18/03/2011 Analytical Model μ i s can be approximated by the departure rate of the WLAN with i saturated stations [11]. Reusing the previous throughput analysis: (12) After computing the state probabilities we obtain the average number of backlogged stations (13) Applying Littles formula, we compute the average delay (14) [11] C. Foh, M. Zukerman, and J. Tantra, A Markovian Framework for Performance Evaluation of IEEE , IEEE Transactions on Wireless Communications, vol. 6, no. 4, pp. 1276–1265, Paul Patras 36

37
18/03/2011 Analytical Model Next, we compute the average collision probability that minimizes the delay (15) The above only depends on λ, μ and T c, which can be easily estimated at the AP CAC-VI aims to drive the collision probability to the target optimal value of (15) Paul Patras 37

38
18/03/2011 CAC-VI Algorithm Similar to CAC, CAC-VI employs a PI controller and consists of: Measuring the p obs resulting from the current CW configuration (observing the retry flags, similar to CAC) Estimating the arrival & departure rates ( λ and μ) and T c obtained by counting the number of frames and computing their average length Computing the new CW configuration and distributing it to the stations We conduct a control-theoretic analysis to configure the controller Paul Patras 38

39
18/03/2011 Performance Evaluation Validation of the analytical model The CW that minimizes delay obtained with our model is close to the optimal value obtained by means of exhaustive search Paul Patras 39

40
18/03/2011 Performance Evaluation Comparison against EDCA & other approaches [4,5,12] [12] A. Nafaa and A. Ksentini, On Sustained QoS Guarantees in Operated IEEE Wireless LANs, IEEE Transactions on Parallel and Distributed Systems, vol. 19, no. 8, pp. 1020–1033, 2008 Paul Patras 40 CAC-VI substantially outperforms previous proposals in terms of average delay

41
18/03/2011 Performance Evaluation Comparison against EDCA & other approaches [4,5,12] [13] ITU-T, Recommendation G.1010: End-user Multimedia QoS Categories Paul Patras 41 With a 150 ms playback deadline CAC-VI provides better QoE [13]

42
18/03/2011 Additional simulations –Validation of the controllers configuration –Total delay, delay distribution –Support for admission control –Mixed traffic scenario –Non-ideal channel conditions –Real users further validate the performance of our proposal Extension of CAC-VI to support graceful degradation of video flows (IEEE TGaa) Performance Evaluation Paul Patras 42

43
18/03/2011 CAC-VI Summary Extended CAC algorithm with the goal of minimizing the average delay –Dynamically adjusts the CW configuration of IEEE stations to improve the QoE of video traffic –Guarantees quick reaction to changes and stable operation Key advantages of CAC-VI over previous proposals [14] –Analytical foundations guarantees optimal operation –Standard compliant, no additional signaling required [14] P. Patras, A. Banchs, and P. Serrano, A Control Theoretic Scheme for Efficient Video Transmission over IEEE e EDCA WLANs, submitted manuscript, November Paul Patras 43

44
18/03/2011 Outline Motivation & Background Centralized Adaptive Control (CAC) Algorithm Distributed Adaptive Control (DAC) Algorithm Experimental Evaluation Conclusions & Future Work Paul Patras 44

45
18/03/2011 Distributed Adaptive Control (DAC) A different approach to performance optimization: –each station independently computes its own configuration by observing the current WLAN conditions –eliminates potential single point of failure problems and the need for additional signaling in non-infrastructure based topologies. –can operate both under infrastructure and ad-hoc mode Objective: adjust the CW min of each station to drive the WLAN to its optimal point of operation Paul Patras 45

46
18/03/2011 Distributed Adaptive Control (DAC) Paul Patras 46 Throughput is maximized with the following optimal value of the collision probability (16) Only driving the WLANs collision probability to the optimal value may lead to different CW settings This can incur fairness problems

47
18/03/2011 Distributed Adaptive Control (DAC) Paul Patras 47 Only driving the WLANs collision probability to the optimal value can incur fairness problems CW 1 CW 2 CW

48
18/03/2011 Distributed Adaptive Control (DAC) Paul Patras 48 Only driving the WLANs collision probability to the optimal value can incur fairness problems CW 1 CW 2 CW

49
18/03/2011 Distributed Adaptive Control (DAC) Paul Patras 49 Only driving the WLANs collision probability to the optimal value can incur fairness problems Nodes should converge to the same CW to ensure fairness CW 1 CW 2 CW

50
18/03/2011 Distributed Adaptive Control (DAC) Paul Patras 50 Throughput is maximized with the following optimal value of the collision probability (16) Objectives: Drive the average probability that a transmission of a station different from i collides as measured by the station i ( p obs,i ) to the optimal value Drive the average probability that a transmission of station i collides as measured by the station ( p own,i ) to the optimal value (17)

51
18/03/2011 Distributed Adaptive Control (DAC) Based on a classical system from multivariable control theory, each station running an independent PI controller For each station i each controller takes as input an error signal e i and gives as output the CW min,i Paul Patras 51

52
18/03/2011 Distributed Adaptive Control (DAC) The choice of the error signal e i is essential to guarantee a set of desired properties under transient conditions: 1.When the collision probability is far from its desired value, e i will be large in order to trigger a quick reaction towards target. 2.When stations do not share bandwidth fairly, the e i should be large in order to achieve a fair bandwidth sharing. 3.In case of congestion, only the saturated stations should increase their CW min,i. Therefore, we define the error signal as follows: (18) Theorem 3 ensures the terms do not cancel each other Paul Patras 52

53
18/03/2011 Distributed Adaptive Control (DAC) Similar to the centralized solution, first term ensure that the collision probability in the network is driven to the optimal value (19) If two stations do not share the bandwidth fairly due to having different CW min,i, the error should be large. (20) Hence, (21) Paul Patras 53

54
18/03/2011 Distributed Adaptive Control (DAC) How to measure p obs,i and p own,i with current devices? Paul Patras 54

55
18/03/2011 Distributed Adaptive Control (DAC) How to measure p obs,i and p own,i with current devices? To estimate p obs,i we examine the retry flag of the overheard transmissions (same method used for CAC) Paul Patras 55

56
18/03/2011 Distributed Adaptive Control (DAC) How to measure p obs,i and p own,i with current devices? To compute p own,i we use the statistics of the wireless interface and record the number of successful ( T ) and failed ( F ) attempts Paul Patras 56

57
18/03/2011 Steady State Analysis Goal: guarantee the existence of a unique stable point of operation for all stations. By Theorem 3, we prove that in steady state (22) and all stations have the same transmission probability (23) Since (24) To maximize the total throughput, we fix p opt to the optimal value of our previous analysis Paul Patras 57

58
18/03/2011 Stability Analysis Goal: Compute the configuration of the PI controllers We characterize the WLAN with a transfer function H that takes CW min as input and returns E as output We use multivariable control theory tools to analyze the stability of the system Theorem 4 gives the necessary conditions the controllers configuration must fulfil to guarantee stability Employ ZN rules to find the right tradeoff between speed of reaction to changes and oscillations under transients (25) Paul Patras 58

59
18/03/2011 Performance Evaluation Saturated scenario DAC maximizes the total throughput, without requiring to estimate the number of stations and avoiding non-standard mechanisms Paul Patras 59

60
18/03/2011 Performance Evaluation Mixed scenario DAC performs better than any other approach when saturated and non- saturated stations coexist WLAN, as it –minimizes the delay performance of non-saturated stations –does not affect the total throughput nor the delay of the saturated STAs Paul Patras 60

61
18/03/2011 Performance Evaluation Convergence Both existing & new stations converge quickly to the same CW value Paul Patras 61

62
18/03/2011 Performance Evaluation Additional results further validate the performance of DAC –Delay under non-saturation conditions –Mixed unbalance traffic conditions –Stability and speed of reaction to changes –Fairness evaluation Paul Patras 62

63
18/03/2011 DAC Summary Distributed approach to the optimal configuration of IEEE WLANs Relies on foundations from the multivariable control theory field - each station implements a PI controller, which uses only locally available information Succeeds in maximizing throughput performance Provides non-saturated stations with lower delay Provides convergence and stability guarantees Is compliant with the IEEE standard [15] [15] P. Patras, A. Banchs, P. Serrano, and A. Azcorra, A Control Theoretic Approach to Distributed Optimal Configuration of WLANs, IEEE Transactions on Mobile Computing, published online, December Paul Patras 63

64
18/03/2011 Outline Motivation & Background Centralized Adaptive Control (CAC) Algorithm Distributed Adaptive Control (DAC) Algorithm Experimental Evaluation Hardware & Software Platform Implementation Details Performance Evaluation Conclusions & Future Work Paul Patras 64

65
18/03/2011 Experimental Evaluation Few of the previously proposed schemes have been developed in practice –Idle Sense [7] entails significant complexity and involves firmware modifications We show that CAC & DAC can be implemented with COTS hardware and open source drivers without changes We thoroughly evaluate their performance in a medium- scale testbed under a wide set of network conditions Discuss the suitability of CAC & DAC algorithms for real deployments Paul Patras 65

66
18/03/2011 Hardware & Software Platform Soekris net embedded computers Atheros AR5414 based a/b/g wireless cards Gentoo Linux OS, kernel MadWifi open-source driver (v0.9.4 modified): Paul Patras 66 –Enabled dynamic setting of EDCA parameters for best-effort AC –Enabled stations to apply the locally computed EDCA configuration for the case of DAC

67
18/03/2011 Implementation Overview CAC & DAC run as user space applications and communicate with the driver through IOCTL calls Exploit the MadWifi capability of supporting multiple virtual devices with a single physical interface Paul Patras 67

68
18/03/2011 Implementation Details Estimation of p obs –Virtual device in monitor mode with promiscuous configuration –Device configured to pass the frames with link headers –Raw socket used to sniff frames and inspect their retry flag Estimation of p own –Perform a SIOCGATHSTATS IOCTL requests to retrieve detailed information about the transmitted frames –Compute the number of successful and failed TX attempts Successes = ast_tx _packets-ast_tx_xretriesast_tx_noack Failures = ast_tx_longretryast_tx_xretries·MAX_RETRY Paul Patras 68

69
18/03/2011 Implementation Details Contention window update –Compute error signal with the estimated probabilities –Every beacon interval compute the new CW min –Impose lower/upper bounds (default min. (16) and max (1024) values of IEEE a) –Pass the new CW value to the driver as the integer exponent of the nearest power of 2 CW[t] = (int) rint(log 2 ( CW min [t] )) –Commit the new value by issuing the corresponding private IOCTL request Paul Patras 69 (26)

70
18/03/2011 Deployed Testbed Paul Patras 70

71
18/03/2011 Deployed Testbed Paul Patras 71

72
18/03/2011 Deployed Testbed Paul Patras 72 RSSI = -20 dBm RSSI = - 62 dBm

73
18/03/2011 Deployed Testbed Paul Patras 73 RSSI = - 62 dBm

74
18/03/2011 Deployed Testbed Paul Patras 74 Capture effect

75
18/03/2011 Performance Evaluation Operation of CAC The announced CW oscillates between two values, the powers of two closest to the optimal CW min p obs fluctuates stably around the desired p opt Paul Patras 75

76
18/03/2011 Performance Evaluation Operation of DAC DAC drives the collision probability in the WLAN to the desired value Paul Patras 76

77
18/03/2011 Performance Evaluation Operation of DAC Stations operate at different CW min (stations that are closer to the AP and benefit from the capture effect observe a smaller collision rate) Paul Patras 77 E[CW min ] = 92 E[CW min ] = 300 E[CW min ] = 64

78
18/03/2011 Performance Evaluation Throughput performance The use of DAC and CAC improves performance by approximately 19% and 21%, respectively EDCA and DAC suffer from fairness issues EDCA favours stations close to the AP, DAC favour stations far from the AP; CAC is fair Paul Patras 78

79
18/03/2011 Performance Evaluation Dynamic traffic conditions: varying the number of stations CAC DAC CAC & DAC immediately detect the change, the system moves to the optimal point of operation and stations are provided with the new CW DAC requires additional time to reach the desired point of operation Paul Patras 79

80
18/03/2011 Performance Evaluation Impact of hidden nodes Node 3 AP Node 2 TX Power = 8 dBm Node 8 TX Power = 5 dBm Paul Patras 80 (AP)

81
18/03/2011 Performance Evaluation Impact of hidden nodes Node 8 silent Node 2 24 Mbps Node 2 achieves an average throughput of 17.4 Mbps Paul Patras 81 (AP)

82
18/03/2011 Performance Evaluation Impact of hidden nodes Node 2 silent Node 8 24 Mbps Node 8 achieves an average throughput of 17.3 Mbps Paul Patras 82 (AP)

83
18/03/2011 Performance Evaluation Impact of hidden nodes Both nodes transmitting 24Mbps Nodes still have a good link to the AP but cannot hear each other Each achieves approx. 1.5 Mbps Paul Patras 83 (AP)

84
18/03/2011 Performance Evaluation Impact of hidden nodes DAC does not provide appreciable advantages over EDCA CAC sees a dramatic throughput increase ( > 3 times the throughput attained with EDCA) leverages the hidden node problem Paul Patras 84 (AP)

85
18/03/2011 Performance Evaluation Additional experiments conducted for further validation of the implemented prototypes: Different network sizes Variable traffic load Quasi-homogeneous link qualities Assessment of resource consumption Paul Patras 85

86
18/03/2011 Experimental Evaluation Summary Implemented CAC & DAC with COTS hardware Provided detailed description of the implementation of the proposed mechanisms Gave insights on the differences between the theoretical design and practical implementations of the algorithms Conducted exhaustive experiments in a medium-scale testbed, to evaluate their performance under non-ideal channel effects and different traffic conditions Identified those scenarios where a network deployment can benefit from using such adaptive algorithms [16] [16] P. Serrano, P. Patras, V. Mancuso, A. Mannocci, and A. Banchs, Adaptive Mechanisms for Performance Optimization of IEEE WLANs: Implementation Experiences and Practical Evaluation, submitted manuscript, December Paul Patras 86

87
18/03/2011 Outline Motivation & Background Centralized Adaptive Control (CAC) Algorithm Distributed Adaptive Control (DAC) Algorithm Experimental Evaluation Conclusions & Future Work Paul Patras 87

88
18/03/2011 Conclusions Proposed novel centralized & distributed adaptive algorithms –Based on analytical models of the IEEE protocol behavior –Sustained by mathematical foundations from single-/multi-variable control theory, which guarantee their stability and convergence –Substantially outperform both the standard IEEE mechanism and previous adaptive proposals Centralized Adaptive Control (CAC) algorithm –By adjusting the CW it maximizes the total throughput –Provides stations with very good fairness –Alleviates the effects of hidden node topologies –Extended with the goal of minimizing the average delay of video Paul Patras 88

89
18/03/2011 Conclusions Distributed Adaptive Control (DAC) algorithm –Each station implements a PI controller, which uses only locally available information –Maximizes the throughput performance of the WLAN –Provides stations not contributing to congestion with better delay performance Implementation using COTS devices and open-source drivers and experimental evaluation in a real testbed –The proposals can be executed by current devices without any modifications into their hardware and/or firmware –Gained insights on their performance under channel impairments and suitability for real deployments Paul Patras 89

90
18/03/2011 Future Work Undertake a distributed approach the performance optimization of real-time traffic Investigate performance under asymmetric (TCP) traffic Further validate the proposals in a large scale deployment with real users Paul Patras 90

91
Control-Theoretic Adaptive Mechanisms for Performance Optimization of IEEE WLANs: Design, Implementation and Experimental Evaluation Author: Paul Horaţiu Pătraş Advisor: Dr. Albert Banchs Roca Thank you for your attention

92
18/03/2011 Backup Slides Paul Patras 92

93
18/03/2011 IEEE EDCA Configurable parameters: AIFS, CW min, CW max, TXOP Statically set for each AC, independent of network conditions, e.g. number of stations, traffic patterns, etc. Paul Patras 93

94
18/03/2011 Related Work Analytical models of DCF/EDCA performance –Either address the saturation scenario or study non-saturation with specific arrival /source types EDCA configuration proposals –Restricted to specific traffic / based on heuristics Centralized approaches –Lack analytical support / do not consider traffic dynamics Distributed approaches –Involve modifications of the existing hardware/firmware Implementation experiences and experimental studies –Subject to HW changes / limited to simple network conditions Paul Patras 94

95
18/03/2011 Computing p obs using the retry flag To compute p obs we inspect the retry flag of the frames observed by the station on the WLAN Paul Patras 95

96
18/03/2011 Non-saturated scenario CAC significantly outperforms the static configuration, which falsely assumes saturation conditions and uses overly large CW values (CAC) Performance Evaluation Paul Patras 96

97
18/03/2011 Impact of channel errors The impact of channel errors (which could be wrongly interpreted as collisions) is minimal for FER values up to 10% (CAC) Performance Evaluation Paul Patras 97

98
18/03/2011 Delay performance Performance of CAC-VI is almost identical to optimal CW configuration obtained with numerical search CAC-VI minimizes delay (CAC-VI) Performance Evaluation Paul Patras 98

99
18/03/2011 Validation of the designed controller The proposed configuration guarantees stable operation, while being able to rapidly track the changes in the WLAN (CAC-VI) Performance Evaluation Paul Patras 99

100
18/03/2011 (CAC-VI) Performance Evaluation Impact of TXOP setting Stations are provided with almost the same average delay performance, while the standard deviation is kept small Paul Patras 100

101
18/03/2011 (CAC-VI) Performance Evaluation Graceful degradation of video quality The mechanism prevents high priority frames from being discarded even in case of large traffic loads Paul Patras 101

102
18/03/2011 (DAC) Stability Analysis Goal: Compute the configuration of the PI controller {K p, K i } First, we express the relationship between CW min and E Paul Patras 102

103
18/03/2011 (DAC) Stability Analysis Paul Patras 103

104
18/03/2011 (DAC) Stability Analysis We characterize the WLAN with a transfer function H that takes CW min as input and returns E as output E can be computed from the CW min,i s by expressing p own,i and p obs,i as a function of the τ i s and The above gives a non-linear relationship between E and CW min we linearize it when the system suffers small perturbations around its stable point of operation Paul Patras 104

105
18/03/2011 (DAC) Stability Analysis The perturbations suffered by E can be approximated by Where By evaluating the partial derivatives around the stable point of operation: Paul Patras 105

106
18/03/2011 (DAC) Stability Analysis With the system fully characterized by the matrices H and C, the remaining challenge is the setting of {K p, K i } From multivariable control theory, we can prove that the system is guaranteed to be stable if In addition to stability, we aim to find the right tradeoff between speed of reaction to changes and oscillations under transient conditions Ziegler-Nichols rules Paul Patras 106

107
18/03/2011 (DAC) Performance Evaluation Non-saturated scenario In addition to maximizing the saturation throughput, DAC also minimizes the average delay under non-saturation Paul Patras 107

108
18/03/2011 (DAC) Performance Evaluation Stability and speed of reaction to changes The proposed {K p, K i } setting provides a good tradeoff between stability and speed of reaction (with a larger setting the system suffers from instability / with a smaller one it reacts slowly to changes) Paul Patras 108

109
18/03/2011 Performance Evaluation Quasi-homogeneous link qualities A careful power setting aids EDCA and DAC in terms of fairness EDCA yields 10% less total throughput (as station benefit less from the capture effect and thus collisions have a greater impact) Paul Patras 109

110
18/03/2011 Experimental Evaluation Impact of link qualities - Fairness heterogeneous links quasi-homogeneous links A careful setting of the TX power improves fairness Paul Patras 110

111
18/03/2011 Performance Evaluation Dynamic traffic conditions: varying the traffic load Stations initiate 5 successive file transfer to the AP with random idle periods uniformly distributed in the [0,20] seconds interval CAC significantly improves the average transfer time DAC slightly outperforms EDCA, but exhibits larger standard deviation Paul Patras 111

112
18/03/2011 Experimental Evaluation Impact of network size DAC and CAC are able to optimize throughput performance by adjusting the CW to the number of stations present in the WLAN EDCA and DAC incur fairness issues once heterogeneous link conditions appear Paul Patras 112

113
18/03/2011 Experimental Evaluation Resource consumption –low speed CPUs (233 MHz) –reduced physical memory (128 MB) CAC is suitable for commercial APs DAC will not affect the performance of current portable devices which employ faster CPUs Performance can be optimized if implemented at driver level Paul Patras 113 AP (CAC)STA (DAC) CPU time39%28% Physical Memory 1.6 %4.3%

Similar presentations

© 2017 SlidePlayer.com Inc.

All rights reserved.

Ads by Google