Control Mechanisms for Video Streaming Wireless Links

1

Control Mechanisms for Video Streaming Wireless

Links

Athina MarkopoulouElectrical Engineering Dept.

Stanford University

2

Real-Time Multimedia over Packet Networks

Characteristics • Continuous Stream Playout, Real-Time

src

rcv

ideally

t

t

src

rcv

network

t

t

delay

loss jitter

Requirements• low loss, delay, delay jitter

3

Multimedia - Networking

Problems at the interface between multimedia applications and underlying network

Control mechanisms • in the network and/or• at the end-systems

Challenges depend on underlying network(s)

Network

Application

4

The Bigger Picture

wireless

backbone

2

1

5

Media Streaming over Wireless Last-Hop

Streaming to• laptops, cellphones, PDAs, wireless TV

displays

Challenges• limited resources, time variation• strict application requirements

Wireless

ServerMobile

TerminalAccess Point, Server/Proxy

Wireline

6

Example

Tx

Rx

Play

IdeallyTx

Rx

Play

Over wireless

Tx

Rx

Play

Control at the RxTx

Rx

Play

Control at the Tx

7

Problem Statement

Scenario• pre-stored media content at Tx.• interference i, according to a Markov chain with • deliver and play entire content

Objective • maximize the playout quality• minimize the power cost

N i

Tx Rx

rp

8

System State and Controls

(p,r) = system controls in current time slot

(n,i,b, r’) = system state in current time slotn = remaining packets at Txi = channel interference b = available packets at Rxr’ = playout rate in previous slot

pn i b

Tx Rx

r

s(p,i)

9

System Controls at Tx

Control p: transmission power in current slot• s(p,i) : probability of successful reception

• Power Cost: Φ = p

– battery lifetime– interference stress

0 2 4 6 8 100

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

SIR p/i

s(p

,i)

pn i

Tx s(p,i)

b

Rx

r

10

Dynamic Programming Formulation

Define to be the minimum expected cost-to-go

(n,j,b-r;r)

(1-s(p,i))qij

(n,i,b;r’)

(n-1,j,b+1-r;r)

s(p,i)qij

(p,r)

System Evolution

… until n=0

from n=N….

Power Quality

·W

·W +

11

Computing the Optimal Control

A stationary optimal solution (p*,r*) exists and can be obtained by value iteration

Optimal policy:• table p*(n,i,b;r’) and r*(n,i,b;r’)• obtain offline & store in lookup table

12

Special Cases: Individual Controls

Similar formulations – obtain optimal policies Compare: no control, special cases, joint control

Scenario Control at Tx Control at RxNo control fix p fix r=R

Power Control adapt p fix r=R

Power Control & Re-buffering

adapt p r in {0, R}

Playout Control fix p r in {0, r1, r2…R}

Re-buffering fix p r in {0, R}

Joint Control adapt p r in {0, r1, r2…R}

13

Power-Quality Tradeoff (1)

Better Performance

No control

14

Power-Quality Tradeoff (2)

No control

Playout Only

15

Power-Quality Tradeoff (3)

No control

Playout Only

Power OnlyPower+Rebuffering

16

Power-Quality Tradeoff (4)

Power OnlyPower+Rebuffering

No control

Playout Only

Joint control

17

Heuristics

Why heuristics? Justified vs. ad-hoc heuristics

• mimic properties of optimal control

Steps• Power-only heuristic • Playout-only heuristic• Joint power-playout heuristic

18

Power Heuristic

in

Tx s(p,i)

b

Rx

rFix playout r=R

Optimal power:

Backlog pressure X(n,i,b) • has structural properties:

Heuristic: approximate • mimicking those properties

X̂b

X

n=1

n=N

p

19

Playout - today

Purpose: choose r(b)

Fixed Threshold Heuristic • L thresholds for buffer occupancy b•

pn

Tx s(p,i)

b

Rx

ri

b

r

r=rl

Bl Bl+1

r=R

r=0BL

B2 B1

r=0r=R/2r=R

fix p

s(i)

20

Playout Heuristic

Observation: channel not taken into account yet

Adaptive Threshold Heuristic• adapt rate • and adapt thresholds

b

Rx

riB2 B1

r=0r=R/2r=Rs(i)

21

Joint Power-Playout Heuristic

i bp

n

Tx Rx

r

s(p,i)

Tx side: • compute X(n,b)• compute power p:

Rx side: • estimate i, compute p and s(p,i)• adjust thresholds, compute playout r• feedback to Tx

^

22

Joint heuristic performs well

Joint OptimalJoint Heuristic

Optimal Playout

Optimal Power+Rebuffering

23

Demo: no-control vs. joint heuristic

For the same interference scenario

For the same power consumption

Compare the playout quality

original

no control

Joint heuristic

24

Comparison Details Controls off Joint Heuristic

25

Wireless Video - Summary Contributions

• Joint power-playout control• Modeled in a dynamic programming framework• Developed simple, efficient heuristics

Extensions • Additional Channels and Responsive Interference • Additional Controls• Content-Aware Control • Apply to protocols (802.11h)

“Joint Power-Playout Control Schemes for Media Streaming over Wireless Links”, in IEEE Packet Video 2004, Markopoulou joint work with Y.Li, N.Bambos, J.Apostolopoulos

26

Extension: adding more controls

Additional Controls:• Tx: control scheduling

– how many units to transmit and which to drop

• Rx: motion-aware playout– slowdown video scenes with low or no motion

Results:• trade-off: playout speed variation vs. distortion• effect of playout variation is less perceived

R(t)bn

Tx Rx

control scheduling

content-aware playout (r)

“Joint Packet Scheduling and Content-Aware Playout Control for Media Streaming over Wireless”, invited paper in IEEE MMSP 2005, A. Markopoulou joint work with Y.Li, N.Bambos, J.Apostolopoulos

27

Example of Motion-Aware Playout

Motion-unaware playout

Motion-aware playout

28

Future Directions Multimedia over IP

• Cross-layer optimization• Content distribution

Network Dependability • From traditional QoS to Reliability & Security• Measurements for diagnosis and control

Network shared by independent selfish entities• Network-adaptive applications• How bad is selfish routing? • Selfishness in other contexts?

Interaction …

29

Appendix

30

Responsive Interference - Setup

Primary Media link, background PCMA links Pairs of Tx-Rx randomly chosen from area

(500x500 wrapped in a torus) Background: geometric durations, Bernoulli

arrivals Free space path loss G~1/d^4, noise

1^(12) Estimate I using previous timeslot N=100, initial 5slots, buffer B=10

Heuristic gains: 60% in power, 66% in QoS

31

Responsive Interference- Power

32

Responsive Interference- Quality

33

Adding mode control: (p, m, r)

Add a control m:• Packets transmitted in a time slot

Add a cost Psi(m) Modify Bellman equations

34

Power-only heuristic Fix playout r=R and find optimal power

p*• similarly to [B&R (1997), B&K(2000), B&Li(2004)]

• where

Heuristic: approximate X, plug it in p.

soft backoff

hard backoff

aggressive

X

i

p*

35

[Performance evaluation cont’d]

Simulated other channels Simulated responsive interference Found low sensitivity to r-

parameters