ee 579: wireless and mobile networks design & laboratory
TRANSCRIPT
![Page 1: EE 579: Wireless and Mobile Networks Design & Laboratory](https://reader030.vdocument.in/reader030/viewer/2022012520/61948d830a25c65fe64b645b/html5/thumbnails/1.jpg)
EE 579: Wireless and Mobile Networks Design & Laboratory
Lecture 4
Amitabha Ghosh Department of Electrical Engineering
USC, Spring 2014
Lecture notes and course design based upon prior semesters taught by Bhaskar Krishnamachari and Murali Annavaram.
![Page 2: EE 579: Wireless and Mobile Networks Design & Laboratory](https://reader030.vdocument.in/reader030/viewer/2022012520/61948d830a25c65fe64b645b/html5/thumbnails/2.jpg)
2
Outline
¾ Administrative Stuff
¾ Presentation by Professor Kyle Konis ([email protected])
¾ Lab Assignment 1
¾ Video over Wireless
![Page 3: EE 579: Wireless and Mobile Networks Design & Laboratory](https://reader030.vdocument.in/reader030/viewer/2022012520/61948d830a25c65fe64b645b/html5/thumbnails/3.jpg)
3
Motivation ¾ Mobile Video Traffic Projection
q Over 66% of all mobile data traffic will be from video by 2017 q 7.4 exabytes (EB) out of 11.2 EB (1 EB = 1018 bytes)
![Page 4: EE 579: Wireless and Mobile Networks Design & Laboratory](https://reader030.vdocument.in/reader030/viewer/2022012520/61948d830a25c65fe64b645b/html5/thumbnails/4.jpg)
4
Motivation ¾ Evolution of Smart Devices
q 8.6 billion handheld mobile devices and 1.2 billion M2M by 2017 q 2.7 GB/month by 2017, as compared to 342 MB/month in 2012
![Page 5: EE 579: Wireless and Mobile Networks Design & Laboratory](https://reader030.vdocument.in/reader030/viewer/2022012520/61948d830a25c65fe64b645b/html5/thumbnails/5.jpg)
5
Motivation ¾ 2G to 3G to 4G and Beyond
q Higher bandwidth, lower latency, increased security q 4G (2012): only 0.9% connections, but 14% of mobile data traffic q 4G (2017): only 10% connections, but 45% of total traffic
![Page 6: EE 579: Wireless and Mobile Networks Design & Laboratory](https://reader030.vdocument.in/reader030/viewer/2022012520/61948d830a25c65fe64b645b/html5/thumbnails/6.jpg)
6
Content-Pipe Divide ¾ Content Providers
q Media companies, end-users, operators of CDN and P2P
q Generate content treating the network as simply a means for communication (dumb pipes)
Transcode Generate multimedia
Frames Shaping Queuing Marking Dropping
Transportation network
DIVIDE
¾ Pipe Providers q ISPs, equipment & network
management vendors, municipalities
q Treat every content equally as simply bits to be transported between nodes (dumb content)
![Page 7: EE 579: Wireless and Mobile Networks Design & Laboratory](https://reader030.vdocument.in/reader030/viewer/2022012520/61948d830a25c65fe64b645b/html5/thumbnails/7.jpg)
7
Content Aware Networking ¾ Protocol Fairness
q Rate fair: Each flow gets half the capacity
q Rate-Distortion fair: Flow1 gets more
¾ A New Protocol Design Paradigm q Utilize content characteristics q Allocate resources based on the optimality criteria that are
reflective of the content q More adaptive and effective network protocols that are rate-
distortion fair
![Page 8: EE 579: Wireless and Mobile Networks Design & Laboratory](https://reader030.vdocument.in/reader030/viewer/2022012520/61948d830a25c65fe64b645b/html5/thumbnails/8.jpg)
8
Kartik Pandit, Amitabha Ghosh, Dipak Ghosal, and Mung Chiang, "Content Aware Optimization for Video Delivery over WCDMA," EURASIP Journal on Wireless Communications and Networking, July 2012. URL: http://anrg.usc.edu/~amitabhg/papers/EURASIP-2012.pdf
Content Aware Video Delivery over 3G WCDMA Networks
![Page 9: EE 579: Wireless and Mobile Networks Design & Laboratory](https://reader030.vdocument.in/reader030/viewer/2022012520/61948d830a25c65fe64b645b/html5/thumbnails/9.jpg)
9
Network Model ¾ Cellular Uplink
q Increasing demand for high data rate ú EVDO RA (1.8 Mbps), LTE (50 Mbps)
q A single WCDMA cell, with a base station serving all users
q Each user transmits a pre-encoded video upstream
q Videos are encoded as GOP (Group of Pictures) structures
¾ Degrees of Freedom – Control q Scheduling (send or drop frame) q Transmission power
![Page 10: EE 579: Wireless and Mobile Networks Design & Laboratory](https://reader030.vdocument.in/reader030/viewer/2022012520/61948d830a25c65fe64b645b/html5/thumbnails/10.jpg)
10
Video Model ¾ Group of Pictures (GOP)
q Successive frames organized into a repetitive structure ú I frame (intra) – coded independently ú P frame (predictive) – motion-compensated difference, depends on
previous P frame ú B frame (bipredictive) – depends on previous and following P/I frames
q Idea: Drop unimportant frames without hurting the quality
B
PI P
B BBGOP: IPBBPBB
Directed acyclic graph Arrows indicate dependency
![Page 11: EE 579: Wireless and Mobile Networks Design & Laboratory](https://reader030.vdocument.in/reader030/viewer/2022012520/61948d830a25c65fe64b645b/html5/thumbnails/11.jpg)
11
Video Model ¾ Scalable Video Coding (H.264)
q Base Layer q Enhancement Layers q Each layer requires more resources q Temporal, Spatial, and Quality scalability
![Page 12: EE 579: Wireless and Mobile Networks Design & Laboratory](https://reader030.vdocument.in/reader030/viewer/2022012520/61948d830a25c65fe64b645b/html5/thumbnails/12.jpg)
12
Problem Formulation ¾ Use optimization theory to allocate resources
q Rate
¾ What do we want to optimize? (Objective function) q Some measure of video quality (e.g., PSNR, distortion)
¾ What are the constraints? q Interference (or SINR) q Available power
¾ What are the variables? q Transmit power q Scheduling decision
User i frame j
Binary variable: 0 if frame j of user i is transmitted; 1 if dropped
![Page 13: EE 579: Wireless and Mobile Networks Design & Laboratory](https://reader030.vdocument.in/reader030/viewer/2022012520/61948d830a25c65fe64b645b/html5/thumbnails/13.jpg)
13
Objective ¾ Maximize PSNR / Minimize Distortion
q PSNR: An objective metric q Expressed in decibel (dB) q Good values > 20-30 (range: 0-100)
Total distortion per GOP:
![Page 14: EE 579: Wireless and Mobile Networks Design & Laboratory](https://reader030.vdocument.in/reader030/viewer/2022012520/61948d830a25c65fe64b645b/html5/thumbnails/14.jpg)
14
Constraints ¾ SINR – signal to
interference plus noise ratio
¾ Achievable rate
User i’
User i
![Page 15: EE 579: Wireless and Mobile Networks Design & Laboratory](https://reader030.vdocument.in/reader030/viewer/2022012520/61948d830a25c65fe64b645b/html5/thumbnails/15.jpg)
15
Constraints ¾ Set of dropped frames for user i
¾ Required rate to transmit the selected frames
¾ Achievable rate under SINR
¾ Constraint: Required rate should be <= achievable rate
![Page 16: EE 579: Wireless and Mobile Networks Design & Laboratory](https://reader030.vdocument.in/reader030/viewer/2022012520/61948d830a25c65fe64b645b/html5/thumbnails/16.jpg)
16
Optimization Formulation ¾ Content-Aware Distortion-Fair Optimization (CADF)
q Minimize the sum of distortions over a GOP for all videos subject to SINR constraints
q An NP-hard problem (MINLP) q Can solve efficiently using heuristics
![Page 17: EE 579: Wireless and Mobile Networks Design & Laboratory](https://reader030.vdocument.in/reader030/viewer/2022012520/61948d830a25c65fe64b645b/html5/thumbnails/17.jpg)
17
A Sample Result ¾ Frame-level distortion: Comparison of CADF scheme with
Foschini-Miljanic scheme
![Page 18: EE 579: Wireless and Mobile Networks Design & Laboratory](https://reader030.vdocument.in/reader030/viewer/2022012520/61948d830a25c65fe64b645b/html5/thumbnails/18.jpg)
18
Jiasi Chen, Amitabha Ghosh, Josphat Magutt, and Mung Chiang, "QAVA: Quota Aware Video Adaptation," ACM CoNEXT, pp. 121--132, Nice, France, December 2012. URL: http://anrg.usc.edu/~amitabhg/papers/CoNEXT-2012.pdf
QAVA: Quota Aware Video Adaptation
![Page 19: EE 579: Wireless and Mobile Networks Design & Laboratory](https://reader030.vdocument.in/reader030/viewer/2022012520/61948d830a25c65fe64b645b/html5/thumbnails/19.jpg)
19
Motivation: The Conflict ¾ Emerging Trends
q Video traffic becoming dominant (>66% by 2017) q Usage-based pricing becoming prevalent
ú AT&T wireless (Jan 2012): $30/$50 for 3/5 GB (baseline) + $10 per GB
ú Verizon Wireless (July 2011): $30/$50/$80 for 2/5/10 GB (baseline) + $10 per GB
¾ Can the user consume more content without worrying about
the wallet?
¾ Is every bit needed for everyone at all times?
![Page 20: EE 579: Wireless and Mobile Networks Design & Laboratory](https://reader030.vdocument.in/reader030/viewer/2022012520/61948d830a25c65fe64b645b/html5/thumbnails/20.jpg)
20
QAVA: Graceful Tunable Tradeoff
Distortion
Cost
Videos watched
Cost
Within budget
Distortion
Minimize
A 3-way tradeoff
# Videos watched
Supply
Size of the video (bit-rate)
Video compressibility Usage profile
![Page 21: EE 579: Wireless and Mobile Networks Design & Laboratory](https://reader030.vdocument.in/reader030/viewer/2022012520/61948d830a25c65fe64b645b/html5/thumbnails/21.jpg)
21
Modular Architecture ¾ Three Modules
q Video Profiler ú Exploit video compressibility
from motion vectors
q User Profiler
ú Predict user’s future data consumption from past history
q Stream Selector ú Choose the right bitrate to
maximize video quality subject to budget
Video Profiler (offline)
motion vectors, bitrates
utility (MOS, PSNR)
User Profiler (online)
past data consumption
predicted consumption
Stream Selector (online)
video request
bit rate video
![Page 22: EE 579: Wireless and Mobile Networks Design & Laboratory](https://reader030.vdocument.in/reader030/viewer/2022012520/61948d830a25c65fe64b645b/html5/thumbnails/22.jpg)
22
Modular Architecture
Adaptively choose the right bit rates
User Profiler (online)
Stream Selector (online)
Video Delivery at right bit rate
Video Profiler (offline)
Video Request
User device Content provider’s server
Access Network
Backbone
![Page 23: EE 579: Wireless and Mobile Networks Design & Laboratory](https://reader030.vdocument.in/reader030/viewer/2022012520/61948d830a25c65fe64b645b/html5/thumbnails/23.jpg)
23
Online Stream Selection: An Example Budget = 3 Goal: Maximize total utility (video quality)
(1,1) v11
v12
video1
(2,2)
(2,1) v21
v22
video2
(4,2)
Online Greedy: v12, v21 Total utility: 2+2 = 4 Total cost: 2+1 = 3
Offline Optimal: v11, v22 Total utility: 1+4 = 5 Total cost: 1+2 = 3
(utility, cost)
![Page 24: EE 579: Wireless and Mobile Networks Design & Laboratory](https://reader030.vdocument.in/reader030/viewer/2022012520/61948d830a25c65fe64b645b/html5/thumbnails/24.jpg)
24
Problem Formulation
Maximize the sum of utilities of all the selected videos, subject to q Exactly one version of each request is granted q Total cost of all the selected versions must be within budget
Online Multi-Choice Knapsack Problem
Budget
# of videos requested # of versions of video i Utility of version j of video i Cost of version j of video i 1 if version j of video i is selected; 0 otherwise
![Page 25: EE 579: Wireless and Mobile Networks Design & Laboratory](https://reader030.vdocument.in/reader030/viewer/2022012520/61948d830a25c65fe64b645b/html5/thumbnails/25.jpg)
25
Evaluation: Video Profiler from MOS ¾ Videos
q 20 diverse H.264 clips q Resolution 640 x 480 q Duration 20 sec q Each video encoded at 100, 150, 200, 300, Kbps
¾ Shown to 20 participants on iPhone4 held at ~50 cm
¾ Participants rated in 1-5 MOS scale q 1: very good (imperceptible distortion) q 5: very annoying
MOS: Mean Opinion Score (subjective video quality metric)
![Page 26: EE 579: Wireless and Mobile Networks Design & Laboratory](https://reader030.vdocument.in/reader030/viewer/2022012520/61948d830a25c65fe64b645b/html5/thumbnails/26.jpg)
26
Results
100 150 200 250 3000
0.5
1
1.5
2
2.5
3
Bitrate
Mea
n Pe
rcep
tual
Dis
torti
on
Low discriminationModerate discriminationHigh discrimination
100 150 200 250 3000
0.5
1
1.5
2
2.5
3
3.5
4
BitrateM
ean
Perc
eptu
al D
isto
rtion
SportsLandscapeMusicTalkshowTV/movieCartoon
MOS for different types of users (consistency across user behavior)
MOS for different types of videos (consistency across motion vector, distortion, rating)
Bitrate"M
ean
Perc
eptu
al D
isto
rtion"
![Page 27: EE 579: Wireless and Mobile Networks Design & Laboratory](https://reader030.vdocument.in/reader030/viewer/2022012520/61948d830a25c65fe64b645b/html5/thumbnails/27.jpg)
27
Results: Overall QAVA Benefit ¾ 1430 video requests randomly generated over 30 days
¾ Video duration normally distributed with mean 30 sec and s.d. 5 sec
QAVA user can watch all videos at low budget Benefit of QAVA decreases for sufficiently large budget Non-QAVA user cannot watch all videos below 11 GB quota
6 8 10 12 14 161
1.5
2
2.5
3
3.5
4
4.5
Cost (GB)
Aver
age
Dis
torti
on
QAVA (SS2)QAVA (SS1)Without QAVAHindsight optimal
Infesible for non-QAVAusers
![Page 28: EE 579: Wireless and Mobile Networks Design & Laboratory](https://reader030.vdocument.in/reader030/viewer/2022012520/61948d830a25c65fe64b645b/html5/thumbnails/28.jpg)
28
Implementation
Implementation on Android
Goals
q Understand consumption behavior of real people q Understand user-perception of video quality q Evaluate the algorithm q Fun to run a trial involving real people
![Page 29: EE 579: Wireless and Mobile Networks Design & Laboratory](https://reader030.vdocument.in/reader030/viewer/2022012520/61948d830a25c65fe64b645b/html5/thumbnails/29.jpg)
29
Princeton Trial
¾ Set Up q 15 volunteers with Android phones q ~500 videos encoded at 25 Kbps granularity (100 Kbps – 500 Kbps)
Database logs: q Video request q Time stamp q User ID / Android ID q MB of video delivered
Video request User and video info request
Tomcat webserver (QAVA server)
MySQL DB
![Page 30: EE 579: Wireless and Mobile Networks Design & Laboratory](https://reader030.vdocument.in/reader030/viewer/2022012520/61948d830a25c65fe64b645b/html5/thumbnails/30.jpg)
30
Android App Screenshots