Authors: HUAHUI WU, MARK CLAYPOOL, and ROBERT KINICKI Presented By Siddharth Singla Jangsung Lee Adjusting Forward Error Correction with Temporal Scaling for TCP-Friendly Streaming MPEG

OVERVIEW Introduction Introduction Background Background Analytical Model Analytical Model Analytic Experiments Analytic Experiments Simulation Experiments Simulation Experiments Conclusion Conclusion

Introduction Volume of data traversing the Net increasing quickly Volume of data traversing the Net increasing quickly Overcoming short-term congestion & avoiding collapse handled by TCP Overcoming short-term congestion & avoiding collapse handled by TCP UDP preferred over TCP for multimedia applications UDP preferred over TCP for multimedia applications Suggestions to make all applications TCP- friendly Suggestions to make all applications TCP- friendly New TCP-friendly protocols for streaming proposed New TCP-friendly protocols for streaming proposed => Congestion! However, Fortunately, Therefore,

Introduction Router AQM techniques respond efficiently to congestion Router AQM techniques respond efficiently to congestion Media Scaling Media Scaling Temporal scaling Temporal scaling Randomly dropping packets can be degraded quality Randomly dropping packets can be degraded quality 3% packet loss is converted into 30% frame loss 3% packet loss is converted into 30% frame loss FEC makes the receiver detect and correct errors by adding redundant data FEC makes the receiver detect and correct errors by adding redundant data Retransmission is not feasible for multimedia Retransmission is not feasible for multimedia

Introduction FEC + TCP friendly protocol is proposed FEC + TCP friendly protocol is proposed However, FEC itself can cause data rate to be reduced However, FEC itself can cause data rate to be reduced

Enhancements in this article Use currently static FEC or adapt FEC to packet loss without regard to TCP friendly data rate constraints Use currently static FEC or adapt FEC to packet loss without regard to TCP friendly data rate constraints Model to account for temporal scaling to preserve real-time video playout Model to account for temporal scaling to preserve real-time video playout Compare new analytic model with simulations Compare new analytic model with simulations Introduction

OVERVIEW Introduction Introduction Background Background

Background 1. TCP friendly flows TCP friendly flows: bandwidth usage in steady-state is no more than a TCP flow under similar network conditions TCP friendly flows: bandwidth usage in steady-state is no more than a TCP flow under similar network conditions s: packet size s: packet size p: packet loss probability p: packet loss probability tRTT: round-trip time tRTT: round-trip time tRTO: retransmit timeout tRTO: retransmit timeout

Background 1. TCP friendly flows T provides upper bound on TCP friendly sending rate T provides upper bound on TCP friendly sending rate Prevents flows from consuming more than fair share of BW Prevents flows from consuming more than fair share of BW * * PADHYE, J., FIROIU, V., TOWSLEY, D., AND KURO SE, J Modeling TCP throughput: A simple mod el and its empirical validation

2. Forward Error Correction (FEC) Add (N −K) redundant packets to the K original packets and send N packets as one frame Add (N −K) redundant packets to the K original packets and send N packets as one frame Frame is successfully reconstructed if >= K correct packets are received Frame is successfully reconstructed if >= K correct packets are received FEC can cause additional delay FEC can cause additional delay Software FEC can be done in real-time with data rates up to 100 Mbps or use H/W to speedup FEC encoding Software FEC can be done in real-time with data rates up to 100 Mbps or use H/W to speedup FEC encoding Background

2. Forward Error Correction (FEC) Probability of K packets being transmitted successfully Probability of K packets being transmitted successfully This equation ignores bursty nature of packets This equation ignores bursty nature of packets Background

3. MPEG Use both intraframe and interframe Use both intraframe and interframe compression compression I frames focus on encoding similarities within a video scene I frames focus on encoding similarities within a video scene P frames are encoded based on motion P frames are encoded based on motion differences from preceding P frames differences from preceding P frames B frames are encoded based motion differences from preceding and succee ding P frames B frames are encoded based motion differences from preceding and succee ding P frames MPEG repeats a GOP MPEG repeats a GOP Background

3. MPEG I frame is more important than P and P is more important than B I frame is more important than P and P is more important than B Background

4. Temporal Scaling Application rate adjust to available bitrate Application rate adjust to available bitrate Low priority frames are discarded before tx Low priority frames are discarded before tx Strategies for temporal scaling: Strategies for temporal scaling: B frames are discarded evenly before discarding I or P frame B frames are discarded evenly before discarding I or P frame P frames are discarded from the back P frames are discarded from the back (last) to the front of the GOP pattern (last) to the front of the GOP pattern I frames are never discarded I frames are never discarded Background

Background 4. Temporal Scaling

OVERVIEW Introduction Introduction Background Background Analytical Model Analytical Model

Analytical Model Determine playable frame rate of TCP- friendly video flows with adjusted FEC and temporal scaling in the presence of network packet loss Determine playable frame rate of TCP- friendly video flows with adjusted FEC and temporal scaling in the presence of network packet loss Also provide probability of successful transmission and playout for each MPEG frame type Also provide probability of successful transmission and playout for each MPEG frame type Derive formulas for transmission rate and playable frame rate Derive formulas for transmission rate and playable frame rate

Analytical Model 1. Software Layers and Parameters RF : maximum playable frame rate RF : maximum playable frame rate SI, SP, SB: number of packets for eac h I, P, or B frame SI, SP, SB: number of packets for eac h I, P, or B frame NP, NB: number of P or B frames in on e GOP NP, NB: number of P or B frames in on e GOP

Analytical Model NPD, NBD: number of P or B frames sent per GOP after temporal scaling NPD, NBD: number of P or B frames sent per GOP after temporal scaling SIF, SPF, SBF: number of FEC packets added to each I, P, or B frame SIF, SPF, SBF: number of FEC packets added to each I, P, or B frame s: packet size (in bytes) s: packet size (in bytes) p: packet loss probability p: packet loss probability tRTT: round-trip time (in milliseconds) tRTT: round-trip time (in milliseconds)

2. Successful Frame Transmission Probabilities Successful transmission of each frame type: Successful transmission of each frame type: qI = q(SI + SIF, SI, p) qI = q(SI + SIF, SI, p) qP = q(SP + SPF, SP, p) qP = q(SP + SPF, SP, p) qB = q(SB + SBF, SB, p) qB = q(SB + SBF, SB, p) Analytical Model

3. Playable Frame Rate GOP rate should be constant for real time playout and is given by GOP rate should be constant for real time playout and is given by Rf: the target full-motion frame rate Rf: the target full-motion frame rate Temporal scaling maintains constant GOP rate Temporal scaling maintains constant GOP rate Playable Rate of I Frames RI is given by Playable Rate of I Frames RI is given by RI = G · qI · DI RI = G · qI · DI Analytical Model

3. Playable Frame Rate Playable Rate of P Frames with tempor al scaling is given by: is given by Playable Rate of P Frames with tempor al scaling is given by: is given by Analytical Model

3. Playable Frame Rate Playable Rate of B Frames with tempor al scaling is given by: Playable Rate of B Frames with tempor al scaling is given by: Analytical Model

3. Playable Frame Rate Total Playable Frame Rate is the sum of playable frame rates for each frame type Total Playable Frame Rate is the sum of playable frame rates for each frame type R = RI + RP + RB R = RI + RP + RB If no frame is discarded due to scaling then R If no frame is discarded due to scaling then R Analytical Model

3. Playable Frame Rate

Optimal Playable Frame Rate: Optimal Playable Frame Rate: Varies with temporal scaling and amount of FEC used Varies with temporal scaling and amount of FEC used Optimum playable frame rate R under TCP-friendly rate constraint can be derived from Optimum playable frame rate R under TCP-friendly rate constraint can be derived from Analytical Model

OVERVIEW Introduction Introduction Background Background Analytical Model Analytical Model Analytic Experiments Analytic Experiments

Analytic Experiments Analyze performance temporally-scaled MPEG video without FEC, with fixed FE C, and with adjusted FEC with TCP friendly constrained bit rates Analyze performance temporally-scaled MPEG video without FEC, with fixed FE C, and with adjusted FEC with TCP friendly constrained bit rates Use 4 distinct repair schemes & MPEG video + FEC are temporally scaled to meet TCP-friendly constraints. Use 4 distinct repair schemes & MPEG video + FEC are temporally scaled to meet TCP-friendly constraints. Fixed FEC (1/0/0) Fixed FEC (1/0/0) Fixed FEC (4/2/1) Fixed FEC (4/2/1) Adjusted FEC Adjusted FEC Non-FEC Non-FEC

Analytic Experiments System Parameter Settings

Playable frame rates of the 4 schemes Analytic Experiments

Adjusting FEC

Analytic Experiments Temporal Scaling Pattern

OVERVIEW Introduction Introduction Background Background Analytical Model Analytical Model Analytic Experiments Analytic Experiments Simulation Experiments Simulation Experiments

Simulation Experiments Compare analytic results with simulation results Compare analytic results with simulation results Analytic experiments assumed a few parameters Analytic experiments assumed a few parameters Simulation is carried out for: Simulation is carried out for: Inaccurate Loss Prediction Inaccurate Loss Prediction Bursty Loss Bursty Loss Variable Round-Trip Times Variable Round-Trip Times Variable MPEG Frame Size Variable MPEG Frame Size

Inaccurate Loss Prediction Simulation Experiments

Bursty Loss Simulation Experiments

Variable Round-Trip Times Simulation Experiments

Variable MPEG Frame Sizes Simulation Experiments

Practical Considerations 3 practical issues are needed to be addressed 3 practical issues are needed to be addressed Dynamically change a GOP based on network conditions Dynamically change a GOP based on network conditions Arbitrarily long GOPs that would make exhaustive search prohibitive Arbitrarily long GOPs that would make exhaustive search prohibitive Case when network and MPEG paramet ers are not known ahead of time Case when network and MPEG paramet ers are not known ahead of time Simulation Experiments

OVERVIEW Introduction Introduction Background Background Analytical Model Analytical Model Analytic Experiments Analytic Experiments Simulation Experiments Simulation Experiments Conclusion Conclusion

Conclusion Propose analytic model to achieve better frame rate based on several parameters Propose analytic model to achieve better frame rate based on several parameters Determine optimal FEC and temporal scaling under different network conditions Determine optimal FEC and temporal scaling under different network conditions Compare analytic model with simulation results Compare analytic model with simulation results Experimental results show that the model could be used in practice to improve fram e rate while maintaining TCP friendly BW Experimental results show that the model could be used in practice to improve fram e rate while maintaining TCP friendly BW

THANK YOU