Measurement and Performance Study of PERT for On-demand Video - - PowerPoint PPT Presentation

Introduction Background Experiment Summary Back-up

Measurement and Performance Study of PERT for On-demand Video Streaming

Bin Qian A.L.Narasimha Reddy

Department of ECE Texas A&M University

PFLDNeT 2010

Introduction Background Experiment Summary Back-up

Outline

1

Introduction

2

Background

3

Experiment NS2 Simulation Linux Test

4

Summary

Introduction Background Experiment Summary Back-up

Motivation Current TCP is not suitable for video streaming applications. In the Internet, many other services (HTTP , FTP , P2P) compete for bandwidth.

Introduction Background Experiment Summary Back-up

Related Work . . . Boyden et al, 2007 TCP can function adequately with a 1.5 higher bandwidth than required stream rate in unconstrained streaming. Wang et al, 2008 TCP generally provides good streaming performance when the achievable TCP throughput is roughly twice the media bitrate, with only a few seconds of startup delay.

Introduction Background Experiment Summary Back-up

Problem How well can TCP support streaming, when T/µ ≤ 2.0? T is the achievable TCP throughput. µ is the video playback bitrate.

Introduction Background Experiment Summary Back-up

Previous Work . . . PERT = Probabilistic Early Response TCP Sumitha et al, 2007 explored the performance of PERT in homogeneous environment. Kiran et al, 2008 made PERT adaptive to heterogeneous environments.

Introduction Background Experiment Summary Back-up

Probabilistic Early Response PERT learns about network congestion by measuring delay

Introduction Background Experiment Summary Back-up

Window Adjustment Mechanism ... Aggressive Window Increasing W = W + α α ≥ 1

Introduction Background Experiment Summary Back-up

Window Adjustment Mechanism ... 3 modes Tcompete = 0.65 * maximum queuing delay When T < Tmin, high-speed mode When T > Tcompete, TCP-compete mode When Tmin < T < Tcompete, safe mode

Introduction Background Experiment Summary Back-up

Window Adjustment Mechanism ... High-speed mode α = αmax = 32 TCP-compete mode α = 1 + p′/p

p′ is the early response probability p is the congestion loss probability

Safe mode α = αmin = 1

Introduction Background Experiment Summary Back-up

Window Adjustment Mechanism Conservative Window Decreasing W = W × (1 − β) β = q′/(q′ + q)

q′ is the estimated queuing delay q is the maximum queuing delay

so W ≥ W/2

Introduction Background Experiment Summary Back-up

Queuing Behavior PERT enqueues more packet earlier and less later ...

200 400 600 800 1000 1200

Queue Position

1 10 100 1000 10000 100000

Frequency

Frequency vs. Queue Position

PERT TCP

Introduction Background Experiment Summary Back-up NS2 Simulation

Setup

Introduction Background Experiment Summary Back-up NS2 Simulation

Parameters Exploration

21 23 26 30 34

CBR Streams Number

1.0 1.2 1.4 1.6 1.8 2.0 2.2 2.4

T/u T/u vs. CBR Streams Number

21 23 26 30 34

CBR Streams Number

5.0 7.5 10.0 12.5 15.0 17.5 20.0 22.5 25.0 27.5 30.0 32.5

Bandwidth (Mbits) Bandwidth vs. CBR Streams Number

CBR FTP HTTP

Introduction Background Experiment Summary Back-up NS2 Simulation

Performance Metric CBR stream is successful if fraction of late packets < 10−4 Video streaming quality is evaluated by fraction of successful CBR streams

Introduction Background Experiment Summary Back-up NS2 Simulation

Simulation Results . . . In low range [1.0-1.4], it drops drastically as T/µ decreases In high range [1.4-2.0], it changes slightly as T/µ increases

1.0-1.2 1.2-1.4 1.4-1.6 1.6-1.8 1.8-2.0

T/u (Start-up Delay 10 secs)

10 20 30 40 50 60 70 80 90 100 110

Fraction of Successful CBR Streams (%)

1000 1000 1000 1000 1000 200 200 200 200 200

Fraction of Successful CBR Streams vs. T/u

PERT RENO CUBIC

Introduction Background Experiment Summary Back-up NS2 Simulation

Simulation Results . . . PERT > RENO and CUBIC in T/µ range [1.0 - 1.4]

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29

Start-up Delay (secs) ( T/u 1.0-1.2, Loss Rate 0.056 )

5 10 15 20 25

Fraction of Successful CBR streams (%) Fraction of Successful CBR streams vs. Start-up Delay RENO PERT CUBIC

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29

Start-up Delay (secs) ( T/u 1.2-1.4, Loss Rate 0.045 )

20 40 60 80 100 120

Fraction of Successful CBR streams (%) Fraction of Successful CBR streams vs. Start-up Delay RENO PERT CUBIC

Introduction Background Experiment Summary Back-up NS2 Simulation

Simulation Results . . . PERT > RENO & PERT ≈ CUBIC in T/µ range [1.4 - 1.8]

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29

Start-up Delay (secs) ( T/u 1.4-1.6, Loss Rate 0.034 )

50 60 70 80 90 100 110

Fraction of Successful CBR streams (%) Fraction of Successful CBR streams vs. Start-up Delay RENO PERT CUBIC

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29

Start-up Delay (secs) ( T/u 1.6-1.8, Loss Rate 0.030 )

70 75 80 85 90 95 100 105

Fraction of Successful CBR streams (%) Fraction of Successful CBR streams vs. Start-up Delay RENO PERT CUBIC

Introduction Background Experiment Summary Back-up NS2 Simulation

Simulation Results PERT > RENO and CUBIC in loss rate range [0.02 - 0.06]

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29

Start-up Delay (secs) ( Loss Rate 0.02-0.04, T/u 1.68)

60 65 70 75 80 85 90 95 100

Fraction of Successful CBR streams (%) Fraction of Successful CBR streams vs. Start-up Delay RENO PERT CUBIC

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29

Start-up Delay (secs) ( Loss Rate 0.04-0.06, T/u 1.26)

20 30 40 50 60 70 80 90

Fraction of Successful CBR streams (%) Fraction of Successful CBR streams vs. Start-up Delay RENO PERT CUBIC

Introduction Background Experiment Summary Back-up Linux Test

Test Bed Bandwidth 15 Mbps Delay 45 ms Buffer 500 Kb Avatar 1080p HTTP streaming

Introduction Background Experiment Summary Back-up Linux Test

Test Results . . . PERT helps to reduce the playback glitches TCP Variants PERT RENO CUBIC Late Picture Skipping # 5.5 33.5 30.5 Audio Output Starving # 3.0 11.0 7.5

Introduction Background Experiment Summary Back-up Linux Test

Test Results PERT responses early before packet loss. PERT adjusts the window smoothly.

1000 2000 3000 4000 5000 6000 7000 8000 9000

Time (0.01s)

20 40 60 80 100 120 140 160 180

CWND Size (Mbytes) CWND Size vs. Time

PERT RENO CUBIC

Introduction Background Experiment Summary Back-up

Conclusions PERT and CUBIC push T/µ constraint to roughly 1.4. PERT > RENO, over all T/µs, loss rates and start-up delays. PERT > CUBIC, over low T/µs, high loss rates and strict start-up delays constraints.

Introduction Background Experiment Summary Back-up

Future Work Carry out more evaluations and comparisons against other protocols. Deploy and measure PERT in error-prone wireless networks.

Introduction Background Experiment Summary Back-up

Thank You !

Introduction Background Experiment Summary Back-up

Probabilistic Early Response Parameters The parameters are currently fixed, and can be chosen adaptively Tmin = 5ms Tmax = 10ms Pmax = 0.05

Introduction Background Experiment Summary Back-up

α adjustment Steady state throughput equations: βPERT(p + p′ − p ∗ p′)/αPERT = βTCP ∗ p/αTCP αTCP = 1 βPERT = βTCP So αPERT = p + p′ − p ∗ p′/p ≈ 1 + p′/p