improving qos and resource utilization for multimedia ... improving qos and resource utilization for...

1

Improving QoS and Resource Utilization for Multimedia Streaming over 3G

Kamal Deep Singh

Co-advisors : David Ros and Laurent ToutainThesis Director : César Viho

2

3G/UMTS Wireless Network

Users of Voice, TCP, …

Video Streaming anywhere in the cell

“Hotspots” with good channel conditions

3

Contributions and Agenda

• Overview

• Video Streaming over 3G

• Header Compression

• Congestion Control for Video Flows

• Conclusion and Future Work

Network

How to improve QoS & resource utilization ?

How can applications adapt to improve the multimedia quality ?

User

4

Agenda

• Overview

• Video Streaming over 3G

• Header Compression

• Congestion Control for Video Flows

• Conclusion and Future Work

Network User

5

Core Network

Overview: 3G/UMTS Network

Radio Access Network

6

Overview: 3G/UMTS Problems

Problems due to the use of IPIP doesn’t support real time streaming requirementsOverhead due to packet header

Problems due to radio conditionsScarce and time varying bandwidthCongestion, wireless losses & large delay

7

Agenda

• Overview

• Video Streaming over 3G

• Header Compression

• Congestion Control for Video Flows

• Conclusion and Future Work

Network User

8

Video Streaming

Video Streaming

• Delay & Jitter• Bandwidth• Packet loss

constraints

9

Video Streaming …

Video Streaming

• Underflow• No Playout

• Packet losses will cause quality distortion

10

Video Streaming …

IP doesn’t support video streaming requirements

Solution is to provide QoS in UMTS network

Different priorities to different flowsService guarantees, …

11

UMTS Downlink High Speed Downlink Packet Access (HSDPA)

Enhanced bit rates: 14.4MbpsFast and adaptive packet scheduling

Packet scheduler looks at a term: channel condition, …

Determines which user will be scheduledThe chances of a user being scheduled

12

Existing HSDPA Schedulers

MAX CI (shown in previous slide)Looks at instantaneous data rate : Unfair

Proportionally Fair (PF)Looks at the ratio of:

QoS SchedulersLook at a “QoS function” and the above ratio

To satisfy QoS constraints or guarantees

Instantaneous data rate

Average throughput

* proportional to channel conditions

*

13

Existing HSDPA Schedulers …

QoS schedulers …QoS function based on the Delta

(Throughput guarantee - Average throughput)

Guarantee not satisfied:Value of QoS function increasesIncreases the probability of being scheduled

Guarantee satisfied:The probability of being scheduled doesn’t increase To be fair to other users

14

Existing HSDPA Schedulers …

How does a QoS function look like ?Rate Guarantee (RG) [Hosein 2002]:

α i.e. a constant in another variant [Lundevall 04]

Bi(t) =

(1 + λi(t) · β · exp (−β · (λi(t)− λ

(i)min)) ∀i ∈ QoS,

1 ∀i ∈ BE,

Guaranteed RateAverage Throughput

Best Effort (BE) users without QoS constraints

15

Existing QoS Schedulers …

RG and its variant have the following problems:

Increase in Best Effort (BE) users deteriorate the QoS

Bias towards certain Rate Guarantees

16

Normalized Rate Guarantee (NRG) Scheduler

Based on the theory in [Hosein 2002] we propose:

Bi(t) =

⎧⎨⎩λ(i)min + λi(t)β · exp

µ−β ·

(λi(t)−λ(i)min)

λ(i)min

¶∀i ∈ QoS,

kBEnBE

∀i ∈ BE.

Normalized with number of BE users

Normalized with Rate Guarantee

Resources allocated to the BE users

17

Simulation Platform

UMTS/HSDPA Simulation Platform

EURANE (NS-2 Extension)

Detailed implementation

SNR traces to simulate physical layer

18

Simulation Platform …H.264 video flows • UM mode

• Per user queues• DiffServ AQM

Pedestrian A, 3km/h

• HSDPA Packet Schedulers• Max. cell capacity: 3.6Mbps

Background TCP flows:• Long lived• WWW

Cell radius: 500m

19

How to evaluate video quality?

Objective evaluationPSNR (peak signal to noise ratio)Pixel by pixel comparison

Subjective evaluationEvaluation done by real humans

20

How to evaluate video quality? …

The importance of subjective quality evaluation:

5/5

Excellent

1/5

Very annoying

3/5

Fair

Scores:

21

Pseudo-Subjective Quality Assessment (PSQA) [Rubino et al. 2002]

Set of Network Parameters

Training

Real Time Usage

Mean Opinion Score

Real Network

RNN

Trained RNN

22

Results4 video users (384kbps, H.264) Qmin: Minimum PSQA score after removing lower 5%NRG performs better for higher loads

Video Users TCP Users

23

Results …

Mix of TCP and WWW trafficNRG performs better for higher loads

Video Users WWW Users

24

Results …Loss Rates for different Rate GuaranteesFor fair comparison

Normalized Best Effort term of RG and RG Lundevall variant

25

Agenda

• Overview

• Video Streaming over 3G

• Header Compression

• Congestion Control for Video Flows

• Conclusion and Future Work

Network User

26

Header Compression for Audio Flows

AudioData

IP/UDP/RTP40 – 120 bytes

100 bytes

AudioData

Compressed Header

2 – 5 bytesSignificant Overhead

Removes redundant information like IP src, dst …

Robust Header Compression (ROHC)

27

ROHC Study: Results

Tradeoff:Compression efficiency Robustness

Propose dynamic negotiationUpdate parameters: error rate, …Efficient when low error Robust when high error

Result: Overall performance improved Lower loss rateBetter compression efficiency

28

ROHC Study: Results …UDP checksum: Corrupted payload droppedCorrupted data may be useful!Studied (UDP-Lite + ROHC) vs. (UDP + ROHC)

ROHC performance doesn't changeProposed CRC strategies

UDP

CRC 50%

CRC 25%

CRC 0%

UDP-Lite

29

Agenda

• Overview

• Video Streaming over 3G

• Header Compression

• Congestion Control for Video Flows

• Conclusion and Future Work

Network User

30

Congestion Control for Video Flows

Motivation• Variable bandwidth, delay …• Congestion, packet losses

How can applications adapt to the network conditions?

31

Existing Schemes

Congestion control for videoTCP: Retransmissions, rate oscillations, …

TFRC [Floyd 2000]Sending Rate is calculated by a TCP model

Better Rate Stability

TCP Friendly

32

Simulation StudyProblem: Rate stabilityRadio conditions TFRC rate variance similar to TCP Improvement with Rate Guarantee (RG) scheduler

33

More Problems: Wireless Losses

Problem: Two types of losses in wireless Networks

Packet drops due to congestionPacket drops due to bad channel conditions

Node B

Router

Wireless loss

Wireless network

IP Packets IP Queue

Congestion loss

34

More Problems …

Inefficiency for TCP, TFRC …Cannot distinguish between these losses.Reduce their sending rate on loss.

How to distinguish Wireless losses from congestion losses?

Previous Work have used Round Trip Time variations

Not reliable

35

Differential dropping in the DiffServ (Green, Yellow & Red)

Video applications mark their packets

Wireless Loss Estimation (Background: DiffServ)

Drop Red packets

Drop Red + Yellow packets

Rarely drop Green packets

I B P B P

Increasing Congestion

36

Wireless Loss Estimation

Wireless Loss Estimation in DiffServ (WLED) Networks:

Red packets are dropped first on congestion

Wireless loss rate (w) is correlated with green loss rate

If loss of yellow packets is not significant

37

WLED:Improves link utilization

There is no change in other properties: TCP friendliness, loss rate, rate stability

But, works only with DiffServ aware applications

WLED: Results

WLED

TFRC

TCP

38

Congestion Control and Adaptive Retransmissions

We improved link utilization in case of wireless losses.

But, lost data still deteriorates the quality!

Solution: We integrate a scheme to retransmit the lost data.

39

Congestion Control and Adaptive Retransmissions …

Retransmission SchemeIf packet has the possibility to arrive before its deadline

No congestionEnables retransmission schemes

CongestionDisable retransmission

Depending on BW retransmit eitherI framesI + P frames orAll frames

WLED scheme integrated

40

Congestion Control and Adaptive Retransmissions: Results

• 10 WWW users in background

41

Agenda

• Overview

• Video Streaming over 3G

• Header Compression

• Congestion Control for Video Flows

• Conclusion and Future Work

Network User

42

Conclusion

Studied QoS scheduling for video streaming over 3G

Subjective quality evaluation

Proposed a new QoS schedulerBetter quality even with high background loadNot biased

Improved header compression performanceUpdating channel states Lower loss rate with variable UDP-Lite checksum coverage

43

Conclusion …

Studied congestion control for video flowsTFRC rate stability not better than TCP as reported in previous works

Proposed a wireless loss estimation scheme

Improved the link utilization

Proposed a retransmission schemeImproved video quality by recovering lost data

44

Results published in following Papers

Kamal Deep Singh and David Ros. ``Normalized Rate Guarantee Schedulerfor High Speed Downlink Packet Access''. In IEEE GLOBECOM 2007, Washington, November 2007.

Kamal Deep Singh, David Ros, Laurent Toutain and César Viho. ``Proportional Resource Partitioning over Shared Wireless Links''. In IEEE 66th VTC, Baltimore, MD, USA, September 2007.

Kamal Deep Singh and David Ros. ``TCP-Friendly Rate Control over High Speed Downlink Packet Access''. In IEEE ISCC'07, Aveiro, Portugal, July 2007.

Kamal Deep Singh, Julio Orozco, David Ros and Gerardo Rubino. ``Streaming of H.264 Video over HSDPA: Impact of MAC-Layer Schedulerson User-Perceived Quality''. In ENST Bretagne Research Report no. RR-2007002-RSM, 2007.

Kamal Deep Singh, David Ros, Laurent Toutain and César Viho. ``Improving Multimedia Streaming over Wireless using End-to-End Estimation of Wireless Losses''. In IEEE 64th VTC, Montreal, Canada, September 2006.

45

Results published in following Papers …

Ana Minaburo, Kamal Deep Singh, Laurent Toutain, Cécile Marc and Elizabeth Martinez. ``Performance Improvement of Multimedia flows by using UDP-Lite and ROHC compression''. In 5th ICICS, Bangkok, Thailand, December 2005.

Ramiro Garcia, Jose Incera and Kamal Deep Singh. ``An Experimental evaluation ofH.264/AVC semantic codec video streaming in a Diffserv network architechture''. In ROC&C 2005, Acapulco, Mexico, November 2005.

Pablo Frank Bolton, Jose Incera and Kamal Deep Singh. ``Objective and subjective characterization of video streams in IP networks''. In ROC&C 2005, Acapulco, Mexico, November 2005.

Ana Minaburo, Kamal Deep Singh, Laurent Toutain and Loutfi Nuaymi. ``Proposedbehavior for robust header compression over a radio link''. In ICC 2004, Paris, France, June 2004.

Submitted

Kamal Deep Singh, Arpad Huszak, David Ros, César Viho and Gabor Jeney. ``Congestion Control and Adaptive Retransmission for Multimedia Streaming overWireless Networks''. Submitted to ICC 2008.

Kamal Deep Singh, Julio Orozco, David Ros and Gerardo Rubino. ``Streaming of H.264 Video over HSDPA: Impact of MAC-Layer Schedulers on User-Perceived Quality''. Submitted to IEEE Transactions.

46

Future Work

Call admission control and management for HSDPA

Real time PSQA feedbackDynamic rate guarantee negotiation

Improve the rate stability of TFRC over HSDPALook at the performance of WLED + retransmission in real networkCongestion control for MPEG4-scalable video codec (SVC)

47

Thank You !

48

Appendix

49

High Speed Downlink Packet Access (HSDPA) Max CI scheduler chooses a user such that :

Instantaneous Data Rate*

*Note that instantaneous data rate is related to the received power

i∗

i∗ = argmaxi{Ri(t)}

50

Existing HSDPA schedulers

Proportionally Fair :

i∗ = argmaxi

½Ri(t)

λi(t)

¾Average Throughput

i∗ = argmaxi

½Bi(t)

Ri(t)

λi(t)

¾Barrier Function

i∗ = argmaxi

½Bi(t)

Ri(t)

λi(t)

¾Barrier Function

i∗ = argmaxi

½Bi(t)

Ri(t)

λi(t)

¾Barrier FunctionQoS schedulers :

51

Existing QoS schedulers

[Hosein 2002] QoS schedulers can be designed based on

Rate Guarantee (RG) [Hosein 2002] uses :

U(λ)

i∗ = argmaxi{Ri(t) · U

0

i (λi(t))}.

U(λ) = log(λ) + 1− exp (−β · (λ− λmin))

Guaranteed Rate

52

Existing QoS schedulers …

RG uses the barrier function :

α i.e. a constant in another variant [Lundevall 04]

Bi(t) =

(1 + λi(t) · β · exp (−β · (λi(t) − λ

(i)min)) ∀i ∈ QoS,

1 ∀i ∈ BE,

Guaranteed Rate

53

Normalized Rate Guarantee (NRG) Scheduling

We propose NRG with the following user utility :

Normalized with number of BE users

Normalized with Rate Guarantee

Resources allocated to the BE users

UQoS(λ) = λmin ·

µlog (λ) +1− exp

µ−β ·

λ−λminλmin

¶¶,

UBE(λ) =kBEnBE

log(λ).

54

Results

4 video users (384kbps, H.264) & 10 TCP long flowsSome users may not get good qualityBetter to drop those users instead of wasting BW

55

Congestion control and Adaptive Retransmissions : Results …

• Wireless loss probability 0.1 and varying WWW users