28
1 Bridging Content-Pipe Divide Amitabha Ghosh Haris Kremo Jiasi Chen Josphat Magutt April 28, 2011

Bridging Content-Pipe Divide

  • Upload
    thora

  • View
    46

  • Download
    0

Embed Size (px)

DESCRIPTION

Bridging Content-Pipe Divide. Amitabha Ghosh Haris Kremo Jiasi Chen Josphat Magutt April 28, 2011. Agenda. Content-Pipe Divide Content-Aware Networking Video Over Wireless Implementation (Theory vs. Practice) Quota-Aware Video Adaptation. Content-Pipe Divide. Content Side - PowerPoint PPT Presentation

Citation preview

Ph.D. Proposal

Bridging Content-Pipe DivideAmitabha GhoshHaris KremoJiasi ChenJosphat Magutt

April 28, 2011

##1AgendaContent-Pipe DivideContent-Aware NetworkingVideo Over WirelessImplementation (Theory vs. Practice)Quota-Aware Video Adaptation#Content-Pipe DivideContent SideMedia companies: own video and musicEnd-users: post video onlineOperators of CDN and P2P systemsSeek the best way to distribute contentThrough multimedia signal processing, caching, relaying, sharing, Treat network as just a means of transportationSeek the best way to manage network infrastructureThrough resource allocation on each link, between links, and end-to-endTreat content as just bits to transport between nodesPipe SideISPsEquipment vendorsNetwork management software vendorsMunicipalities and enterprisesDIVIDE#3Traditional ThinkingSeparation between content generation and transportation

TranscodeGenerate multimediaFramesShapingQueuingMarkingDroppingTransportation networkSeparation#

New ThinkingContent-Aware NetworkingAdjust PHY and MAC layer parameters to suitDrop packets by frame distortion (I, P vs. B)

Network-Aware Content GenerationSVC transcodingJoint summarization + resource allocationGOP: IPBBPBBPBB#Rate-Distortion FairTwo flows competing for BW over a common linkRate Fairness: Each flow gets half the capacityDistortion Fairness: Flow1 gets more capacity than Flow2

Flow1 with less motion helps Flow2 with rich motion#Distortion MetricPSNRCaptures only spatial variation

PCACaptures motion/activity

#Related WorksContent-Aware distortion-Fair dropping [Chiang 09]Minimize max end-to-end distortion in multi-hop wired networksUser-driven, threshold-based dropping based on frame priorities

Discrete time frame selection [Chiang 08]Voice + video, wireless, one-hop, multi-userJARS: Joint Adaptation (summarization), Resource allocation (distributed pricing-based), Scheduling (greedy centralized TDM)

MU-MDP traffic state optimization [van der Schaar 10]Maximize expected discounted accumulated utilityBuffer modeling, value iteration, reinforcement learning, Bellmans equations, stochastic sub-gradient

#Related WorksModulation, MAC retry, path selection [van der Schaar 06]Cross-layer approach to maximize capacity-distortion utilityExhaustive search, greedy algorithm

Rate-distortion optimized streaming [Chou 06]Single user, wired networkScheduling policy vector to minimize expected distortion subject to rate constraint

Media-aware rate allocation [Girod 10]Proxy-server: receiver-driven, proxy-client: sender-drivenPolicy (Markov decision tree): which packets to select for transmissionIterative Sensitivity Analysis (ISA)

#

Problem FormulationCDMA Uplink: An Implementable Solution

: TX power of user i at time t : SINR at BS from user i at time t

Rate:Utility: negative distortion

Goal:

subject to: SINR and deadline constraints

Scheduling vs. Power ControlCSMA vs. CDMA#ImplementationTheory vs. Practice#Closed loop power control for CSMAdriven by video qualityA software defined radio implementation study

Haris Kremo##OutlineImplementation

Power control algorithmtarget received power driven by video qualityrequires video profilingreceived signal strength (RSSI) feedback

Demo setup

Conclusionon theory vs. practice gap#Rice University WARP software defined radio PHY: 802.11 (p-like) OFDM64 carriers across 10 MHztransmit power adjustable in 0.5 dB stepsrange: -20 dBm to 10 dBmBPSK, QPSK, 16-QAM, 64-QAM

MAC: 802.11 DCFcarrier sensing through energy detectionexponential random backoffACK successful reception

programmableXilinx FPGA#Closed loop power controlSelect signal strength at receiver to match desired video qualityAdjust transmit power to achieve that signal strengthPSNR to RSSItargetPSNRreceiver

-

DATAACKcalculate RSSI transmitter ireceiver jtime varying channel#Video profilingTabulate distortion vs. signal strength

Connect transmitter and receiver with a cableFor different fixed power levels in 2dBm steps:stream video and save it on the receiverrecord RSSIcalculate frame-by-frame distortion offline

original videodistortionfixed adjustable powerreceived video

RSSI #Experimental setupFour videos streaming to one receiverHigh Definition (HD) vs. Low Definition (LD)High Motion (HM) vs. Low Motion (LM)

Adjust manually target PSNR

HDHMHDLMLDHMLDLM

#17Theory vs. practiceCDMA vs. CSMAlicensed vs. unlicensed bandconnection based vs. packet based

Hard to calculate video metric in real time

RSSI not a good measure of interference

Practicalities inaccuracies: 1dB resolutionnonlinearities: set power out of rangeoutdated feedback:insufficient packet rate#Quota-Aware Video AdaptationJiasi Chen

April 28, 2011

#1919

System Architecture

End UserEdge ISPInternetContent Providercostdistortion of videos

VideoStores multiple precoded streams of each video#20MotivationWhats the best way to compress videos and stay within budget constraints, while maintaining perceptual quality?#Adaptation EngineAlgorithmInput videoClassifierOutput videoQuotaUser profileVideo profileProfiler

#22User Profiling

#23Optimization ProblemMaximize utilitySubject to budget constraintsSpecial case of knapsack problemOnline algorithm: video requests are not known in advanceAs each request arrives, make an on-the-fly decision of how much to compress#24Divide billing cycle into sessionsIn each session, create a knapsack based on predictionChoose items for knapsack

Optimal to of offline algorithm(Chakrabarty et al., Budget constrained bidding in keyboard auctions and online knapsack problems, Proc. 17th Intl Conf WWW, 2008)Online Algorithm

#25A possible wayOnline Algorithm

#26Bigger legend.

Consumer Cost SavingsDataCostFirst 200 MB$15Each additional 200 MB$15Quota = 200 MB#Rescale, distortion, show different users. Per gigabyte over27Thank you!#