内容案例培训30p06-cache long v1

1、Cache-Induced Opportunistic Cooperative MIMO: A new Paradigm for Content Access NetworksVincent LAU, Director and Founder, Huawei-HKUST Innovation LabDepartment of ECE, Hong Kong University of Science and TechnologyChallenges in Future 5G Wireless Networks25G Wireless Systems 1000X increase in capac

2、ity demand by 2020.a significant portion of the capacity demand comes from high quality video streaming applications.Not only capacity but also stringent real-time QoS demands How to meet the Demands?How to Achieve the Challenging TargetChallenges in 5G Wireless Systems: Wireless Networks are “Inter

3、ference Limited”First-order solutions break the interference barrier!Existing advanced technologies to mitigate interferenceInterference Avoidance:- e.g. TDMA / FDMA / eICIC (ABS) Not efficient. Total network capacity does not scale with K. Interference Averaging:- e.g. CDMANot efficient:- Total net

4、work capacity saturates with SNR (interference limited)Network MIMO:- Cooperative MIMO / Virtual MIMOVery Efficient:- Total capacity scales with K and SNR; Yet, very expensive in terms of backhaul loading.Backhaul can be expensive for dense small cell networks (e.g. China Mobile & SK )3Can we realiz

5、e the full benefits of Cooperative MIMO without expensive backhaul?Coordinated MIMO vs Cooperative MIMONo payload sharing between BS unfavorable topology (interference channel)Payload Sharing between BS favorable topology (broadcast channel)4No data sharing(No Cooperation)with data sharing (MIMO Coo

6、peration)Each BS can only transmit 2 data streams to each mobileEach BS can transmit 4 data streams to the mobileCan we realize the full benefit of MIMO Cooperation without full backhaul?Classical Information Theory: Networks are designed to deliver “random information bits”Performance of networks i

7、s measured by “capacity” = # of random bits/secTopology without backhaul Interference Channels or Relay ChannelsLimited Capacity due to Cut-set boundCapacity with limited backhaulLimited by Cut-set boundsRole of Cache in Wireless Networks:Contents are not “random bits” Contents are “cachable” Cached

8、 Contents may be accessed at a later timeE.g. Video Streaming Cache-Induced MIMO Cooperation (Cache-Induced Topology Change):If the accessed content exists in the BS cache simultaneously, then we can engage in MIMO cooperation without backhaul5Cache-Induced Opportunistic MIMO Cooperation6How to shar

9、e data between transmitters (BSs)?Conventional Method (Cellular Systems)Use BackhaulProposed MethodUse BS-Level CacheChallenge:MIMO Cooperation Gain limited by “how likely requested data are in all the Cache SIMULTANEOUSLY”7Conventional Cache & Proposed Cache are ComplimentaryConventional Caching in

10、 CDN/Internet CoreLoad BalancingProposed Caching for Opportunistic Cooperative MIMO Create MIMO Cooperation Opportunity (without backhaul)Reduce # of hops How is it different from conventional Caching in Internet / CDN?Challenge #1: Cache Design & Data StructureCache Design & MIMO Cooperation Opport

11、unityNave Caching: Probability of MIMO Cooperation Opportunity exponentially small w.r.t. number of Tx9Challenge #2: End-to-End Cross-Layer Design for Video Streaming11Video Streaming QoS Requirements Playback Interruption: it occurs when the playback buffer underflows during the transmission highly

12、 undesirable for the end user experience.Fig. Illustration of a queue trajectory of the playback queue at the mobile user. playback interruption occurs12Video Streaming QoS Requirements Buffer overflow: it causes ing video packets dropped and results in wastage of wireless resource used to transmit

13、these dropped packets.Fig. Illustration of a queue trajectory of the playback queue at the mobile user. buffer overflow occurs13Dynamic Resource ControlTruly Cross-layer Design:Cache Control Adapts to long-term video popularity statisticsMIMO Precoder Adapts to short-term CSI, QSI, Cache-StateJointl

14、y Optimizes the “end-to-end” video streaming performance (playback interruption & buffer overflow)Adaptation to the Channel StatesReveals good “transmission opportunities” w.r.t. fading channelExploit Multi-user Diversity Gain.Adaption to the Queue LengthReveals “urgency of different data flows”.Que

15、stion: How to combine these two different factors into the overall priority?Challenge 1: Requires both Information theory (modeling of the PHY dynamics) & Queueing theory (modeling of the queue dynamics). Challenge 2: complex queue coupling makes the stochastic optimization problem multi-dimensional

16、. Brute-force approach cannot lead to any viable solution. System ModelL Video Files, Size of the l-th video = Fl;2M video streaming users M-antenna BS + M-antenna RS (Cache)Index of video file requested by the k-th userUser Request Profile (URP) Physical Layer ModelMode-0 (Cache State S=0): Current

17、 payload requested by the 2M users are NOT in the RS-Cache.The payload requested can only be served by the BSM users are randomly selected for transmission at the BS (ZFBF over M antennas)Mode-1 (Cache State S=1):The payload requested by the 2M users are in the RS-Cache.All 2M users can be served by

18、 CoMP between the BS and RS (ZFBF over 2M antennas)Received Signal at the k-th user:Equivalent channel gainTransmit power (controlled)Queue Dynamics and QoS MetricQueue Dynamics in Playback BufferControlled Random Arrival PHY Tx Data Rate (Tx power)Playback rate:QJoint Power and Cache Control Policy

19、 and Problem FormulationPower Control PolicyGlobal System State:Video Streaming PerformanceInterruption ProbabilityOverflow ProbabilityAverage Transmit Power CostTwo Timescale Stochastic Optimization ProblemShort Timescale Power Control:Long Timescale Cache ControlCache CostCache Probability Control

20、Low Complexity SolutionInner Power Control SolutionMulti-level Water-Filling Power ControlAdapts to CSI, QSI, Cache StateWater Level depends on QSI via “Priority Function” (Solution of PDE)Due to infinite state spaceOuter Cache ControlChallenge:Objective Function of P0 depends on the optimal value o

21、f the MDP Inner Problem. No closed form characterizationConvex Approximation of P0Stochastic Gradient (learning based) No need to know the popularity statistics of URPSystem PerformanceSystem Performance (Preliminary)Simulation setupA media streaming system with L = 6 media files and K = 4 BS-user p

22、airsBandwidth = 1MhzThe size of each video file is 600M Bytes BaselinesBaseline 1 (CSI-only, no RS): PHY reduces to the Mode 0 ZF. The power control is adaptive to CSI only.Baseline 2 (Q-weighted Sum Rate, no RS): Mode 0 ZF, power control is obtained by solving Q-weighted-sum-rate maximization:Basel

23、ine 3 (Q-weighted DF): DF PHY (the channel between the BS and RS is 20dB larger than direct link), power control is obtained by solving the above Q-weighted-sum-rate maximization25Power and Interruption Probability Tradeoff PerformanceLarge gain over all baselinesgain increases with the cache size P

24、ower and Playback Buffer Overflow Probability Tradeoff PerformanceLarge gain over all baselinesThe overflow probability of baselines increases with transmit power. The proposed is opposite!Computation Complexity and Cache Update LoadingComplexity is similar to CSI only policy and is lower than DF du

25、e to low complexity MDP solution and simple PHY (ZF)Large SNR gain with small backhaul consumptionMuch smaller backhaul loading compared to Conventional CoMPComputation Complexity and Cache Update LoadingConclusionsConclusions30A new paradigm of Content Access Networks by using Cache at BSCache at BS