A Ph D Dissertation Using Bandwidth Estimation to
A Ph. D Dissertation Using Bandwidth Estimation to Optimize Buffer and Rate Selection for Streaming Multimedia over IEEE 802. 11 Wireless Networks Mingzhe Li Committee: Prof. Mark Claypool – CS, Worcester Polytechnic Institute Prof. Robert Kinicki – CS, Worcester Polytechnic Institute Prof. Emmanuel Agu – CS, Worcester Polytechnic Institute Prof. Constantinos Dovrolis - College of Computing, Georgia-Tech Ph. D Dissertation Computer Science Dept. Worcester Polytechnic Institute
Motivation § Improve streaming performance over wireless networks § Increasingly deployment of streaming multimedia over wireless networks – Wireless link capacity increasing – Streaming techniques become maturing – Hardware price decreasing § Streaming techniques are not optimized for wireless networks – Streaming rate selection – Playout buffer – Bandwidth estimation Ph. D Dissertation – 18 February 2021 2
Streaming Rate Selection Goal: Adjust streaming media's data rate according to the network conditions. High Best Network Bottleneck Mid Streaming Client Low Streaming Server Ph. D Dissertation – 18 February 2021 Packet losses 3
Streaming Rate Selection (cont. ) § Traditional approaches – Based on packet loss rate, Round Trip Time (RTT), or application quality metrics (Frame rate/image quality). – Based on packet pair bandwidth estimation § Rate selection in wireless networks – Packet loss could be caused by bit error, but not congestion – Round trip time variation could be caused by MAC layer contention, but not queuing delay – Inaccurate bandwidth estimations § Traditional approach results in selecting a rate higher/lower than the best rate in wireless networks Ph. D Dissertation – 18 February 2021 4
Playout Buffer Goal: Smooth variation in bandwidth and streaming bit rate at the cost of startup delay. Buffer Network Bottleneck Media Streaming Server Streaming Client Select buffer size: § Tradeoff between buffer underflow events and startup delay § Ideally, adaptive to the network conditions Ph. D Dissertation – 18 February 2021 5
Playout Buffer (cont. ) § Traditional approaches – Fixed size for commercial applications – Adaptive based on network measurements § Playout buffer in wireless networks – High variance in available bandwidth – Traditional adaptive buffer size is insufficient § Result in more buffer underflow events in wireless networks or long startup delay Ph. D Dissertation – 18 February 2021 6
Bandwidth Estimation § Traditional approaches – Designed for precisely estimate the bandwidth in wired networks – Converge based on searching algorithms – Provide limited bandwidth information § Impacted by wireless networks – Inaccurate results – Long estimation time – High intrusiveness 7 Ph. D Dissertation – 18 February 2021
Requirements for Bandwidth Estimation for Streaming over Wireless Networks § Adequate accuracy in bandwidth estimation for wireless networks § Fast estimation time § Low intrusiveness § Broader bandwidth information – Rate selection needs available bandwidth – Playout buffer size selection needs variance in bandwidth 8 Ph. D Dissertation – 18 February 2021
Dissertation Block Diagram 9 Ph. D Dissertation – 18 February 2021
Dissertation Components Approach Wireless Bandwidth Estimation Tool (WBest) M A S I P Buffer and Rate Optimization for Streaming (BROS) M A I P Emulated Streaming (Emu. S) M: Analytical Model A: Optimization Algorithm S: Simulation Ph. D Dissertation – 18 February 2021 I P I: Implementation P: Performance Evaluation 10
Outline § Introduction § The Dissertation § Wireless Bandwidth Estimation Tool (WBest) – Related Work – Algorithm – Evaluation § Contributions § Future work 11 Ph. D Dissertation – 18 February 2021
Related Bandwidth Estimation Techniques § Packet dispersion – bprobe [Crovella 1996], pathrate [Dovrolis, 2001], etc – Measure the packet pair/train dispersion to estimate the bottleneck capacity – Fast, medium accuracy, low intrusiveness § Self-loading probe (Probe Rate Model - PRM) – path. Chirp [Ribeiro 2003], pathload [Jain 2003], etc – Vary traffic load and use the packet delay to estimate the available bandwidth at the narrow link – Slow, high accuracy, high intrusiveness § Probe Gap Model (PGM) – Spruce [Strauss 2003], IGI [Hu 2003], etc. – Use the changes of the gap between packets to estimate the crossing traffic, then estimate the available bandwidth with given bottleneck capacity – Assume a known bottleneck capacity – Fast, medium accuracy, low intrusiveness 12 Ph. D Dissertation – 18 February 2021
Packet Dispersion Bottleneck router L: Ci: ∆in: ∆out: Packet size Bottleneck capacity Initial gap Dispersed gap 13 Ph. D Dissertation – 18 February 2021
Example: Packet Dispersion with Wireless Contention Probing traffic Contending traffic / Co-channel interference 14 Ph. D Dissertation – 18 February 2021
Simulation Results of Packet Dispersion in Wireless Networks § Higher and inconstant overhead – Inter frame spaces – CSMA/CA handshakes – Random delay between two back-to-back frames § Physical layer rate adaptation – Multipath fading – Signal attenuation – Bursty errors § MAC layer contention – Shared media – IEEE 802. 11 MAC Retry (ARQ) – Exponential backoffs § High Bit Error Rate (BER) – Lost frames – MAC layer ARQ Ph. D Dissertation – 18 February 2021 NS 2 Simulations (V 2. 27): 1. Ideal channel 2. Fading channel with multirate PHY 3. Contending channel 4. Bit Error Rate (BER) 15
Wireless Bandwidth Estimation Tool (WBest) § Objective: – fast, low intrusiveness, adequately accurate estimation of available bandwidth and variance of bandwidth in wireless networks § Two-step algorithm: – a packet pair technique to estimate the effective capacity of the wireless network; – a packet train technique to estimate the mean and standard deviation of available bandwidth 16 Ph. D Dissertation – 18 February 2021
Terminologies § Effective capacity (Ce) – Indicates the maximum capability of the wireless network to deliver network layer traffic – Includes rate adaptation impact – Includes the BER impact § Available bandwidth (A) – Maximum unused bandwidth – Impacted by contending/crossing traffic – A = Ce – S, where S is bandwidth reduction caused by crossing and contending traffic 17 Ph. D Dissertation – 18 February 2021
WBest Assumptions § Assume the last hop wireless network (hth hop) is the bottleneck link with a single FCFS queue and: § Assume no significant changes in network conditions between the two steps (estimating Ce and A) 18 Ph. D Dissertation – 18 February 2021
Estimating Effective Capacity (Ce) § Send n packet pairs to estimate Ce: – Ti : dispersion time of ith packet pair (seconds) – L : packet size (bytes) § Use the median of n estimations to minimize the impact of crossing and contending traffic 19 Ph. D Dissertation – 18 February 2021
Estimating Available Bandwidth (A) § A packet train of m packets is sent at effective capacity (Ce) to estimate available bandwidth § FCFS queuing at AP – R : dispersion rate S : crossing/contending traffic – S’ : reduced crossing/contending traffic § Estimate the contending and crossing traffic (S) using the dispersion rate (R) Ph. D Dissertation – 18 February 2021 20
Estimating Available Bandwidth (A) (cont. ) § Mean available bandwidth (A) A § Infer the variance of available bandwidth Measured R 21 Ph. D Dissertation – 18 February 2021
Advantages of WBest § Fast estimation – Not based on convergence searching algorithm § Low intrusiveness § Reasonably accurate bandwidth estimation § Reduce the wireless impact on searching algorithm – Delay measurement – Wireless losses § Others – Estimation of effective capacity Ce – Estimation of variance in available bandwidth 22 Ph. D Dissertation – 18 February 2021
Evaluation § § § Build testbed – Open source drivers – Wireless sniffer Various wireless conditions – Traffic load – Power saving mode – Rate adaptation Implementation of WBest Ph. D Dissertation – 18 February 2021 § Compare with: – IGI/PTR v 2. 0 [Hu 2003] (PGM/PRM) – path. Chirp v 2. 4. 1 [Ribeiro 2003] (PRM) – pathload v 1. 3. 2 [Jain 2003] (PRM) 23
Evaluation (cont. ) § 15 evaluation cases – – Various Crossing/Contending traffic TCP and UDP protocols Rate adaptation Power saving mode § Repeat 30 times for each case § Each run consists of four tools and one Constant Bit Rate (CBR) to measure the “ground truth” of available bandwidth § In a lab with radio-shield paint, during middle of night § Performance metrics – Available bandwidth (Mbps) – Estimation time /convergence time (seconds) – Intrusiveness (MBytes) Ph. D Dissertation – 18 February 2021 24
Results of Available Bandwidth (UDP crossing traffic) Ground truth Ground Truth 25 Ph. D Dissertation – 18 February 2021
Results of Estimation Time (UDP crossing traffic) 26 Ph. D Dissertation – 18 February 2021
Results of Intrusiveness (UDP crossing traffic) 27 Ph. D Dissertation – 18 February 2021
Summary of all 15 Cases (Estimation Times) Error is different between estimated bandwidth and the “ground truth” measured by CBR throughput. Ph. D Dissertation – 18 February 2021 28
Summary of all 15 Cases (Intrusiveness) 29 Ph. D Dissertation – 18 February 2021
Outline § Introduction § The Dissertation § Wireless Bandwidth Estimation Tool (WBest) – Related Work – Algorithm – Evaluation § Contributions § Future work 30 Ph. D Dissertation – 18 February 2021
Contributions § Wireless Bandwidth Estimation Tool (WBest) – Fast, low intrusiveness, accurate estimation for IEEE 802. 11 networks – Broader bandwidth information: Effective capacity, Available bandwidth and variance – Implemented and evaluated in Linux § Packet dispersion model in wireless networks – Includes channel rate, contending, BERs, MAC layer retry and exponential backoff – Validated by simulations and measurements 31 Ph. D Dissertation – 18 February 2021
Contributions (cont. ) § Playout buffer model – Markov Chain model based on streaming rate and distribution of available bandwidth – Validated by measurements in wireless networks 32 Ph. D Dissertation – 18 February 2021
Contributions (cont. ) § Buffer and Rate Optimization for Streaming (BROS) – Optimizes streaming rate and initial buffer size based on WBest estimations – Adjustable based on target buffer underflow probability – Implemented and evaluated in wireless networks • Reduce buffer underflow by nearly 100% • Reduce frame loss by nearly 100% • Reduce total buffer delay by 80% § Emulated Streaming (Emu. S) system – Multiple encoded layers and configurable playout buffer – Provides performance information: frame loss, buffer delay and underflow – Integrated with WBest and BROS 33 Ph. D Dissertation – 18 February 2021
Future Work § Extend WBest to other types of wireless networks such as WWANs § Extend BROS/WBest for live or interactive applications § Improve WBest to use streaming multimedia data to estimate available bandwidth § Improve WBest to report loss rate and delay information, which may be used to improve the media repair techniques 34 Ph. D Dissertation – 18 February 2021
Acknowledgements § § § Prof. Claypool and Prof. Kinicki Prof. Agu Prof. Dovrolis from Georgia-Tech Faculty/Staff of Computer Science Dept. , WPI Jae Chung, Feng Li, Rui Lu, Hao Shang, Huahui Wu, and everyone from PEDS and CC groups § Attendees today § My Family 35 Ph. D Dissertation – 18 February 2021
A Ph. D Dissertation Using Bandwidth Estimation to Optimize Buffer and Rate Selection for Streaming Multimedia over IEEE 802. 11 Wireless Networks Mingzhe Li Committee: Prof. Mark Claypool – CS, Worcester Polytechnic Institute Prof. Robert Kinicki – CS, Worcester Polytechnic Institute Prof. Emmanuel Agu – CS, Worcester Polytechnic Institute Prof. Constantinos Dovrolis - College of Computing, Georgia-Tech Ph. D Dissertation Computer Science Dept. Worcester Polytechnic Institute
Rate Selection in IEEE 802. 11 g WLAN 2 -minute Windows Media streaming session Streaming over IEEE 802. 11 g WPI campus WLAN 11 encoding levels Ph. D Dissertation – 18 February 2021 37
Buffer Events in IEEE 802. 11 g WLAN 38 Ph. D Dissertation – 18 February 2021
BROS rate selection 39 Ph. D Dissertation – 18 February 2021
BROS buffer optimization 40 Ph. D Dissertation – 18 February 2021
Buffer Underflow (Rate adaptation case) 41 Ph. D Dissertation – 18 February 2021
Frame Loss (Rate adaptation case) 42 Ph. D Dissertation – 18 February 2021
Total Buffer Delay (Rate adaptation case) 43 Ph. D Dissertation – 18 February 2021
- Slides: 43