際際滷

際際滷Share a Scribd company logo

A Psychophysical Design towards Fair Bandwidth Allocation among VoIP Sessions

?
?
Shannon Chen
Shannon Chen

This document describes research on developing a fair bandwidth allocation mechanism for Voice over IP (VoIP) sessions. The proposed mechanism uses psychophysical modeling and exponential quantization of sending rates to improve fairness and overall user satisfaction. It was evaluated through simulations comparing its performance to naive and Skype-based approaches in sustaining large numbers of simultaneous calls over limited bandwidth. The results demonstrated that the proposed mechanism increases fairness by serving more concurrent users and improves accumulated user satisfaction levels under constrained network conditions.

1 of 37
Network and Systems Laboratory nslab.ee.ntu.edu.tw A Psychophysical Design towards Fair Bandwidth Allocation among VoIP Sessions Chien-nan Chen 秀槻 Network and Systems Laboratory Graduate Institute of Networking and Multimedia National Taiwan University 2012/06/27 Advisors: Polly Huang and Hao-hua Chu Copyright ? 2012 1
Network and Systems Laboratory nslab.ee.ntu.edu.tw Copyright ? 2012 2
Network and Systems Laboratory nslab.ee.ntu.edu.tw Adaptation Psychophysics Sending QoS Rate Bandwidth Measurement Fairness QoE Rate Control Performance Assessment Copyright ? 2012 3
Network and Systems Laboratory nslab.ee.ntu.edu.tw Roadmap Mechanism Simulation ? QoS vs. QoE Design ? Sustainable ? Sending Rate vs. ? Rate control number of users ? Call-based Satisfaction ? Sending rate ? Accumulated simulation quantization satisfaction ? Comparison with Skype Modeling Analysis Copyright ? 2012 4
Network and Systems Laboratory nslab.ee.ntu.edu.tw Roadmap Mechanism Simulation ? QoS vs. QoE Design ? Sustainable ? Sending Rate vs. ? Rate control number of users ? Call-based Satisfaction ? Sending rate ? Accumulated simulation quantization satisfaction ? Comparison with Skype Modeling Analysis Copyright ? 2012 5
Network and Systems Laboratory nslab.ee.ntu.edu.tw Subjecting Codecs AMR-WB SILK ? Widely used in mobile ? The up-to-date codec devices used by Skype ? Nine coding rates ? Variable coding rates (6.6~23.8 kbps) (5.6~40.6 kbps) ? Two sampling rates ? Multiple sampling rates (8 and 16 kHz) (8, 12, 16, 24 kHz) ? ECC embedded, extra ? Wide spectrum of bits for redundancy qualities, extra bits for elaboration of details Copyright ? 2012 6
Network and Systems Laboratory nslab.ee.ntu.edu.tw Copyright ? 2012 ?#?
Network and Systems Laboratory nslab.ee.ntu.edu.tw Copyright ? 2012 ?#?
Network and Systems Laboratory nslab.ee.ntu.edu.tw Roadmap Mechanism Simulation ? QoS vs. QoE Design ? Sustainable ? Sending Rate vs. ? Rate control number of users ? Call-based Satisfaction ? Sending rate ? Accumulated simulation quantization satisfaction ? Comparison with Skype Modeling Analysis Copyright ? 2012 9
Network and Systems Laboratory nslab.ee.ntu.edu.tw Design 6 5 4 MOS 3 2 MOS of fixed-quality tracks 1 ln(br-4.019)+1.515 0 0 5 10 15 20 25 Bitrate (kbps) 30 35 40 45 The exact mathematic model Take only the log property Divide the sending rate into levels with exponential differences Copyright ? 2012 10
Network and Systems Laboratory nslab.ee.ntu.edu.tw Mechanism ? Sending rate is exponentially quantized into levels (which map to equally separated MOSs) ? A call is only allowed to transmit data at one of the level at any time ? Rate is raised to the highest level which the available bandwidth allows ? Rate is dropped to the next lower level when available bandwidth cannot sustain Copyright ? 2012 11
Network and Systems Laboratory nslab.ee.ntu.edu.tw Exponential Quantization (EQ) ? Simple and distributed ? Fairness: increases the number of calls served under the same network capacity ? Performance: increases the accumulated QoE of users served Copyright ? 2012 12
Network and Systems Laboratory nslab.ee.ntu.edu.tw Roadmap Mechanism Simulation ? QoS vs. QoE Design ? Sustainable ? Sending Rate vs. ? Rate control number of users ? Call-based Satisfaction ? Sending rate ? Accumulated simulation quantization satisfaction ? Comparison with Skype Modeling Analysis Copyright ? 2012 13
Network and Systems Laboratory nslab.ee.ntu.edu.tw 1. Fairness: Increase Users Case 1: Available bandwidth increasing (by B) ? Copyright ? 2012 14
Network and Systems Laboratory nslab.ee.ntu.edu.tw 1. Fairness: Increase Users Case 2: Available bandwidth decreasing ? Copyright ? 2012 15
Network and Systems Laboratory nslab.ee.ntu.edu.tw Fairer is Better ? VoIP, like any other interactive networking application, is a multi-party service ? Hoarding resource cannot improve your service quality High Rate Bad Tx Good Tx Quality Quality Good Rx Bad Rx Quality Quality Low Rate Copyright ? 2012 16
Network and Systems Laboratory nslab.ee.ntu.edu.tw 2. Performance: Increase Σ QoE ? Rate change Copyright ? 2012 Normalized by the original rate 17
Network and Systems Laboratory nslab.ee.ntu.edu.tw P-Fair and Accumulated QoE ? Copyright ? 2012 18
Network and Systems Laboratory nslab.ee.ntu.edu.tw Roadmap Mechanism Simulation ? QoS vs. QoE Design ? Sustainable ? Sending Rate vs. ? Rate control number of users ? Call-based Satisfaction ? Sending rate ? Accumulated simulation quantization satisfaction ? Comparison with Skype Modeling Analysis Copyright ? 2012 19
Network and Systems Laboratory nslab.ee.ntu.edu.tw Call-based Simulation ? We simulated 1,000~10,000 simultaneous calls in a backbone link with running background traffic ? The background traffic is adopted from [Fraleigh 03] which suggested a fractional Brownian motion with 124 Mbps average rate ? We simulated an OC-3 backbone link with 155 Mbps capacity Copyright ? 2012 20
Network and Systems Laboratory nslab.ee.ntu.edu.tw Comparison Three scenarios are simulated, where the calls adopt rate adaptation scheme of: 1. Exponential Quantization 2. Na?ve (baseline) Changes in available bandwitdth is evenly distributed to all calls, regardless of their qualities 3. Skype (reality check) By manipulating the bandwidth and recording the resulting rate of Skype, we manage to synthesize adaptation scheme of Skype Copyright ? 2012 21
Network and Systems Laboratory nslab.ee.ntu.edu.tw Number of Calls Served 3000 100% 90% 2500 80% Percentage of call supported Number of supported calls 70% 2000 60% 1500 50% 40% 1000 30% 20% 500 10% 0 0% Number of simultaneous calls Number of simultaneous calls ＊ Exponential Quantization ＊ Na?ve ＊ Skype Copyright ? 2012 22
Network and Systems Laboratory nslab.ee.ntu.edu.tw Accumulated QoE ? Problem: ITU never 8000 define MOS value for a 6000 forced dropped call 4000 Accumulate QoE 2000 ? According to our Na?ve Quant 0 model, MOS -2000 Skype approaches Cinf when -4000 bitrate is zero -6000 Number of simultaneous calls ? Our model outperforms Accumulated QoE when forced drop=-1 others when the dropped calls are given negative scores Copyright ? 2012 23
Network and Systems Laboratory nslab.ee.ntu.edu.tw MOS Distribution Exponential Quantization Naive Skype 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% 1000 3000 5000 7000 9000 1000 3000 5000 7000 9000 1000 3000 5000 7000 9000 Number of simultaneous calls Number of simultaneous calls Number of simultaneous calls Copyright ? 2012 24
Network and Systems Laboratory nslab.ee.ntu.edu.tw Conclusion Aiming at devising a rate control mechanism for VoIP calls, we investigate: ? How users perceive voice quality at different sending rates with two popular speech codecs ? How one allocates the bandwidth such that we gain more users than losing more In result, we develop the simple and distributed EQ scheme that: ? Increase the user population (na?ve 334%; Skype 180%) ? Increase user¨s satisfaction (na?ve +2.3; Skype +1.0) Copyright ? 2012 25
Network and Systems Laboratory nslab.ee.ntu.edu.tw Copyright ? 2012 26
Network and Systems Laboratory nslab.ee.ntu.edu.tw Define MOS of A Track ? Given a quantized-rate track, we now define its QoE. ? For a rate-changing incident as follows: ? MOS of every colored Rate fFLUC blocks is then fFIX weight-averaged by their time durations Time a b c min(a,b) min(b,c) Copyright ? 2012 27
Network and Systems Laboratory nslab.ee.ntu.edu.tw Define MOS of A Track ? For example, MOS of this particular track would be Rate fFLUC r1 f FIX r2 r3 Time a b c ? MOS = {fFIX(r3)*a/2+ fFLUC(r1,r3,a)*a+ fFIX(r1)*(b-a/2-c/2)+fFLUC(r1,r2,c)*c+fFIX(r2)*c/2}/(a+b+c) Copyright ? 2012 28
Network and Systems Laboratory nslab.ee.ntu.edu.tw Proof of Approaching P-Fair ? Copyright ? 2012 29
Network and Systems Laboratory nslab.ee.ntu.edu.tw Proof of P-Fair: Notations ? Copyright ? 2012 30
Network and Systems Laboratory nslab.ee.ntu.edu.tw Properties ? Copyright ? 2012 31
Network and Systems Laboratory nslab.ee.ntu.edu.tw P−K Theorem 1: In time interval [t0, t1], the total number of level-up events P is no less than the number of level-down events K. ? Total amount of bandwidth increased and decreased during [t0,t1] are the same ? By the proof of favoring low calls, more bandwidth is released by a random call decreasing its rate than increasing ? The number of increasing adjustment (P) must be more than the number of increasing adjustment (K) Copyright ? 2012 32
Network and Systems Laboratory nslab.ee.ntu.edu.tw P¨−K¨ ? Copyright ? 2012 33
Network and Systems Laboratory nslab.ee.ntu.edu.tw ? Copyright ? 2012 34
Network and Systems Laboratory nslab.ee.ntu.edu.tw Skype: Fixed Bandwidth Available Sending BW (kbps) Rate (kbps) 70 70 58.08947692 60 65 57.08036923 Sending Rate (kbps) 60 57.98412308 50 55 55.31864615 40 50 50.12886154 30 45 44.36393846 20 40 38.43926154 10 35 34.64332308 0 30 29.42424615 0 10 20 30 40 50 60 70 80 25 26.43064615 Available Bandwidth (kbps) 20 24.02375385 15 21.58646154 10 18.40947692 5 17.77181538 Copyright ? 2012 35
Network and Systems Laboratory nslab.ee.ntu.edu.tw Skype: Bandwidth Change Low BW (kbps) High BW (kbps) Diff (kbps) L->H(kbps/s) H->L (kbps/s) 25 30 5 0.6379381 2.0469237 25 35 10 0.7553864 1.62483514 25 45 20 0.8976954 5.60387293 25 55 30 1.4017246 5.89804405 30 35 5 0.4279915 2.4580033 30 40 10 0.7265226 3.14087255 30 50 20 1.2878194 2.59801165 35 40 5 0.4281306 2.381474 35 45 10 1.0751739 4.16207214 35 55 20 1.6227416 3.86504924 40 45 5 1.1018639 3.50748608 40 50 10 1.932563 2.73371261 40 60 20 2.7938022 2.84590066 45 50 5 0.8966093 1.85362606 45 55 10 0.7646173 2.94752985 50 55 5 0.71419 1.55146863 Copyright ? 2012 36
Network and Systems Laboratory nslab.ee.ntu.edu.tw Granularity 3 Levels 5 Levels 9 Levels 60 50 bandwidth(kbps) 40 30 20 10 0 1 5 9 13 17 1 5 9 13 17 1 5 9 13 17 time(second) time(second) time(second) ＊ Call #1 ＊ Call #2 ＊ Spare Bandwidth Granularity 3 Levels 5 Levels 9 Levels BW Consumption 19.811 kbps 26.780 kbps 29.794 kbps BW Utilization 61.91% 83.69% 93.11% Aggregated QoE 3.107 3.310 3.323 Copyright ? 2012 37

More Related Content

A Psychophysical Design towards Fair Bandwidth Allocation among VoIP Sessions

  • 1. Network and Systems Laboratory nslab.ee.ntu.edu.tw A Psychophysical Design towards Fair Bandwidth Allocation among VoIP Sessions Chien-nan Chen 秀槻 Network and Systems Laboratory Graduate Institute of Networking and Multimedia National Taiwan University 2012/06/27 Advisors: Polly Huang and Hao-hua Chu Copyright ? 2012 1
  • 2. Network and Systems Laboratory nslab.ee.ntu.edu.tw Copyright ? 2012 2
  • 3. Network and Systems Laboratory nslab.ee.ntu.edu.tw Adaptation Psychophysics Sending QoS Rate Bandwidth Measurement Fairness QoE Rate Control Performance Assessment Copyright ? 2012 3
  • 4. Network and Systems Laboratory nslab.ee.ntu.edu.tw Roadmap Mechanism Simulation ? QoS vs. QoE Design ? Sustainable ? Sending Rate vs. ? Rate control number of users ? Call-based Satisfaction ? Sending rate ? Accumulated simulation quantization satisfaction ? Comparison with Skype Modeling Analysis Copyright ? 2012 4
  • 5. Network and Systems Laboratory nslab.ee.ntu.edu.tw Roadmap Mechanism Simulation ? QoS vs. QoE Design ? Sustainable ? Sending Rate vs. ? Rate control number of users ? Call-based Satisfaction ? Sending rate ? Accumulated simulation quantization satisfaction ? Comparison with Skype Modeling Analysis Copyright ? 2012 5
  • 6. Network and Systems Laboratory nslab.ee.ntu.edu.tw Subjecting Codecs AMR-WB SILK ? Widely used in mobile ? The up-to-date codec devices used by Skype ? Nine coding rates ? Variable coding rates (6.6~23.8 kbps) (5.6~40.6 kbps) ? Two sampling rates ? Multiple sampling rates (8 and 16 kHz) (8, 12, 16, 24 kHz) ? ECC embedded, extra ? Wide spectrum of bits for redundancy qualities, extra bits for elaboration of details Copyright ? 2012 6
  • 7. Network and Systems Laboratory nslab.ee.ntu.edu.tw Copyright ? 2012 ?#?
  • 8. Network and Systems Laboratory nslab.ee.ntu.edu.tw Copyright ? 2012 ?#?
  • 9. Network and Systems Laboratory nslab.ee.ntu.edu.tw Roadmap Mechanism Simulation ? QoS vs. QoE Design ? Sustainable ? Sending Rate vs. ? Rate control number of users ? Call-based Satisfaction ? Sending rate ? Accumulated simulation quantization satisfaction ? Comparison with Skype Modeling Analysis Copyright ? 2012 9
  • 10. Network and Systems Laboratory nslab.ee.ntu.edu.tw Design 6 5 4 MOS 3 2 MOS of fixed-quality tracks 1 ln(br-4.019)+1.515 0 0 5 10 15 20 25 Bitrate (kbps) 30 35 40 45 The exact mathematic model Take only the log property Divide the sending rate into levels with exponential differences Copyright ? 2012 10
  • 11. Network and Systems Laboratory nslab.ee.ntu.edu.tw Mechanism ? Sending rate is exponentially quantized into levels (which map to equally separated MOSs) ? A call is only allowed to transmit data at one of the level at any time ? Rate is raised to the highest level which the available bandwidth allows ? Rate is dropped to the next lower level when available bandwidth cannot sustain Copyright ? 2012 11
  • 12. Network and Systems Laboratory nslab.ee.ntu.edu.tw Exponential Quantization (EQ) ? Simple and distributed ? Fairness: increases the number of calls served under the same network capacity ? Performance: increases the accumulated QoE of users served Copyright ? 2012 12
  • 13. Network and Systems Laboratory nslab.ee.ntu.edu.tw Roadmap Mechanism Simulation ? QoS vs. QoE Design ? Sustainable ? Sending Rate vs. ? Rate control number of users ? Call-based Satisfaction ? Sending rate ? Accumulated simulation quantization satisfaction ? Comparison with Skype Modeling Analysis Copyright ? 2012 13
  • 14. Network and Systems Laboratory nslab.ee.ntu.edu.tw 1. Fairness: Increase Users Case 1: Available bandwidth increasing (by B) ? Copyright ? 2012 14
  • 15. Network and Systems Laboratory nslab.ee.ntu.edu.tw 1. Fairness: Increase Users Case 2: Available bandwidth decreasing ? Copyright ? 2012 15
  • 16. Network and Systems Laboratory nslab.ee.ntu.edu.tw Fairer is Better ? VoIP, like any other interactive networking application, is a multi-party service ? Hoarding resource cannot improve your service quality High Rate Bad Tx Good Tx Quality Quality Good Rx Bad Rx Quality Quality Low Rate Copyright ? 2012 16
  • 17. Network and Systems Laboratory nslab.ee.ntu.edu.tw 2. Performance: Increase Σ QoE ? Rate change Copyright ? 2012 Normalized by the original rate 17
  • 18. Network and Systems Laboratory nslab.ee.ntu.edu.tw P-Fair and Accumulated QoE ? Copyright ? 2012 18
  • 19. Network and Systems Laboratory nslab.ee.ntu.edu.tw Roadmap Mechanism Simulation ? QoS vs. QoE Design ? Sustainable ? Sending Rate vs. ? Rate control number of users ? Call-based Satisfaction ? Sending rate ? Accumulated simulation quantization satisfaction ? Comparison with Skype Modeling Analysis Copyright ? 2012 19
  • 20. Network and Systems Laboratory nslab.ee.ntu.edu.tw Call-based Simulation ? We simulated 1,000~10,000 simultaneous calls in a backbone link with running background traffic ? The background traffic is adopted from [Fraleigh 03] which suggested a fractional Brownian motion with 124 Mbps average rate ? We simulated an OC-3 backbone link with 155 Mbps capacity Copyright ? 2012 20
  • 21. Network and Systems Laboratory nslab.ee.ntu.edu.tw Comparison Three scenarios are simulated, where the calls adopt rate adaptation scheme of: 1. Exponential Quantization 2. Na?ve (baseline) Changes in available bandwitdth is evenly distributed to all calls, regardless of their qualities 3. Skype (reality check) By manipulating the bandwidth and recording the resulting rate of Skype, we manage to synthesize adaptation scheme of Skype Copyright ? 2012 21
  • 22. Network and Systems Laboratory nslab.ee.ntu.edu.tw Number of Calls Served 3000 100% 90% 2500 80% Percentage of call supported Number of supported calls 70% 2000 60% 1500 50% 40% 1000 30% 20% 500 10% 0 0% Number of simultaneous calls Number of simultaneous calls ＊ Exponential Quantization ＊ Na?ve ＊ Skype Copyright ? 2012 22
  • 23. Network and Systems Laboratory nslab.ee.ntu.edu.tw Accumulated QoE ? Problem: ITU never 8000 define MOS value for a 6000 forced dropped call 4000 Accumulate QoE 2000 ? According to our Na?ve Quant 0 model, MOS -2000 Skype approaches Cinf when -4000 bitrate is zero -6000 Number of simultaneous calls ? Our model outperforms Accumulated QoE when forced drop=-1 others when the dropped calls are given negative scores Copyright ? 2012 23
  • 24. Network and Systems Laboratory nslab.ee.ntu.edu.tw MOS Distribution Exponential Quantization Naive Skype 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% 1000 3000 5000 7000 9000 1000 3000 5000 7000 9000 1000 3000 5000 7000 9000 Number of simultaneous calls Number of simultaneous calls Number of simultaneous calls Copyright ? 2012 24
  • 25. Network and Systems Laboratory nslab.ee.ntu.edu.tw Conclusion Aiming at devising a rate control mechanism for VoIP calls, we investigate: ? How users perceive voice quality at different sending rates with two popular speech codecs ? How one allocates the bandwidth such that we gain more users than losing more In result, we develop the simple and distributed EQ scheme that: ? Increase the user population (na?ve 334%; Skype 180%) ? Increase user¨s satisfaction (na?ve +2.3; Skype +1.0) Copyright ? 2012 25
  • 26. Network and Systems Laboratory nslab.ee.ntu.edu.tw Copyright ? 2012 26
  • 27. Network and Systems Laboratory nslab.ee.ntu.edu.tw Define MOS of A Track ? Given a quantized-rate track, we now define its QoE. ? For a rate-changing incident as follows: ? MOS of every colored Rate fFLUC blocks is then fFIX weight-averaged by their time durations Time a b c min(a,b) min(b,c) Copyright ? 2012 27
  • 28. Network and Systems Laboratory nslab.ee.ntu.edu.tw Define MOS of A Track ? For example, MOS of this particular track would be Rate fFLUC r1 f FIX r2 r3 Time a b c ? MOS = {fFIX(r3)*a/2+ fFLUC(r1,r3,a)*a+ fFIX(r1)*(b-a/2-c/2)+fFLUC(r1,r2,c)*c+fFIX(r2)*c/2}/(a+b+c) Copyright ? 2012 28
  • 29. Network and Systems Laboratory nslab.ee.ntu.edu.tw Proof of Approaching P-Fair ? Copyright ? 2012 29
  • 30. Network and Systems Laboratory nslab.ee.ntu.edu.tw Proof of P-Fair: Notations ? Copyright ? 2012 30
  • 31. Network and Systems Laboratory nslab.ee.ntu.edu.tw Properties ? Copyright ? 2012 31
  • 32. Network and Systems Laboratory nslab.ee.ntu.edu.tw P−K Theorem 1: In time interval [t0, t1], the total number of level-up events P is no less than the number of level-down events K. ? Total amount of bandwidth increased and decreased during [t0,t1] are the same ? By the proof of favoring low calls, more bandwidth is released by a random call decreasing its rate than increasing ? The number of increasing adjustment (P) must be more than the number of increasing adjustment (K) Copyright ? 2012 32
  • 33. Network and Systems Laboratory nslab.ee.ntu.edu.tw P¨−K¨ ? Copyright ? 2012 33
  • 34. Network and Systems Laboratory nslab.ee.ntu.edu.tw ? Copyright ? 2012 34
  • 35. Network and Systems Laboratory nslab.ee.ntu.edu.tw Skype: Fixed Bandwidth Available Sending BW (kbps) Rate (kbps) 70 70 58.08947692 60 65 57.08036923 Sending Rate (kbps) 60 57.98412308 50 55 55.31864615 40 50 50.12886154 30 45 44.36393846 20 40 38.43926154 10 35 34.64332308 0 30 29.42424615 0 10 20 30 40 50 60 70 80 25 26.43064615 Available Bandwidth (kbps) 20 24.02375385 15 21.58646154 10 18.40947692 5 17.77181538 Copyright ? 2012 35
  • 36. Network and Systems Laboratory nslab.ee.ntu.edu.tw Skype: Bandwidth Change Low BW (kbps) High BW (kbps) Diff (kbps) L->H(kbps/s) H->L (kbps/s) 25 30 5 0.6379381 2.0469237 25 35 10 0.7553864 1.62483514 25 45 20 0.8976954 5.60387293 25 55 30 1.4017246 5.89804405 30 35 5 0.4279915 2.4580033 30 40 10 0.7265226 3.14087255 30 50 20 1.2878194 2.59801165 35 40 5 0.4281306 2.381474 35 45 10 1.0751739 4.16207214 35 55 20 1.6227416 3.86504924 40 45 5 1.1018639 3.50748608 40 50 10 1.932563 2.73371261 40 60 20 2.7938022 2.84590066 45 50 5 0.8966093 1.85362606 45 55 10 0.7646173 2.94752985 50 55 5 0.71419 1.55146863 Copyright ? 2012 36
  • 37. Network and Systems Laboratory nslab.ee.ntu.edu.tw Granularity 3 Levels 5 Levels 9 Levels 60 50 bandwidth(kbps) 40 30 20 10 0 1 5 9 13 17 1 5 9 13 17 1 5 9 13 17 time(second) time(second) time(second) ＊ Call #1 ＊ Call #2 ＊ Spare Bandwidth Granularity 3 Levels 5 Levels 9 Levels BW Consumption 19.811 kbps 26.780 kbps 29.794 kbps BW Utilization 61.91% 83.69% 93.11% Aggregated QoE 3.107 3.310 3.323 Copyright ? 2012 37