CommNet2 & IcoreJoint Workshop on Content Caching & Distributed Storage for Future Communication Networks QoE-Driven Video Streaming and Video Content Caching Xiaohong Peng Adaptive Communications Networks Research Group School of Engineering & Applied Science Aston University Birmingham, UK x-h.peng@aston.ac.uk 1
Outline QoE-Driven Video Streaming QoE-Driven Cache Management 2
Mobile Video Will Generate Over 69% of Mobile Data Traffic by 2018 (Source: CISCO VNI Mobile, 2014) 3
Video Service Connections in LTE 4
A System View of Streaming Protocols Streaming system Internet Encoded video/audio streams (H.264/H.265, VP8/VP9); (MP3) HTTP, RTP, RTSP TCP, UDP IP Bitstream (with MPD) Media Presentation Description Low layer protocols 5
Technical Challenges* Increasing wireless capacity by 1000 times Connecting 20 billion people-oriented devices Connecting 1 trillion objects in the Internet of Things Saving 90% of the energy used Providing latency of under 5 milliseconds (ms) Providing a perceived connection reliability of 99.999% * Defined by 5GPPP-European Commission
600 600 400 400 400 200 200 200 0 0 0 y 600 y y Capacity vs. Coverage -200-200 -200-400 -400-400 -600-600 -600-500 0 x 500 High capacity over a small area -500 0 x 500-500 0 x 500 Lower capacity with wider coverage How to balance between them?
Performance vs. Fairness A llo cated d ata (M b /s) 7 x 105 6 5 4 3 2 1 PI PI(modified) MaxMin propfair bestcqi 0-10 -5 0 5 10 15 20 Average SINR (db) Jane fairness index Jane index over the accumulated allocated capacity, VehA, 1.4MHz, nue=10 0.7 0.65 0.6 0.55 0.5 0.45 0.4 0.35 0.3 0.25 Round Robin maxci PI l, PI h = 0.2, 0.5 PI l, PI h = 0.2, 0.7 PI l, PI h = 0.2, 0.9 0.2 0 100 200 300 400 500 600 700 800 900 1000 time (TTI) High throughput is always companied by low fairness How to balance between them?
Fidelity Vs. Continuity How to balance between them?
It s a Quality Issue QoS QoE QoL Quality of Service Quality of Experience Quality of Life
What is Quality of Experience? The overall acceptability of an application or service, as perceived subjectively by the enduser. Definition of Quality of Experience (QoE), document COM 12- LS 62-E, ITU-T, Jan. 2007, pp. 12. 11
Fundamental Trade-Off Quality Requirement How to maintain a good march between QoS and QoE? 12
A Solution: Adaptive Streaming MPEG-DASH (Dynamic Adaptive Streaming over HTTP) It Is an adaptive bitrate streaming technique: DASH uses existing Internet infrastructure (HTTP web server) to deliver multimedia content to Internet connected devices. Video content is broken into a sequence of small HTTP-based file segments, which could be encoded at different bit rates and clients automatically select the rate of the next segment to download and play back based on current network conditions. MPEG-DASH is for high quality streaming of media content over the Internet, while 3GP-DASH is for adaptive HTTP streaming in wireless mobile networks. Sources: MPEG ratifies its draft standard for DASH". MPEG. 2011-12-02. Retrieved 2012-08-26. ISO/IEC DIS 23009-1:2014 Dynamic adaptive streaming over HTTP (DASH) ETSI 3GPP TS 26.247; Transparent end-to-end packet-switched streaming service (PSS); Progressive Download and Dynamic Adaptive Streaming over HTTP (3GP-DASH) 13
Multiple Bitrate Streams for a Single Video In DASH, a single stream is transcoded into multiple streams with varied bitrates and levels of fidelity quality, such as used in YouTube: 3000Kbps 6000Kbps (1080p) 1500Kbps 4000Kbps (720p) 500Kbps 2000Kbps (480p) 400Kbps 1000Kbps (360) 300Kbps 700Kbps (240) HD ED SD 14
Challenges Remained Recent experimental analysis has concluded: The existing adaptive HTTP streaming players are still at their infancy. The technology is new and it is still not clear how to design an effective rate-adaptation logic for a complex and demanding video streaming application that has to function on top of a complex transport protocol (TCP). (Source: S. Akhshabi, et al., An experimental evaluation of rate-adaptation algorithms in adaptive streaming over http, ACM MMSys, 2011) When should the rate be changed, and by how much? 15
Conventional QoE Assessment Methods QoE Subjective Objective Mean Opinion Score (MOS) Full-reference metrics No-reference metrics Reduced-reference metrics Data metrics Picture metrics Bitstream metrics 16
Objective Metrics for QoEAssessment of Video Streaming QoE (objective) Fidelity based Continuity based PSNR SSIM Blurriness Blockiness Initial buffering delay Rebuffering frequency (pause frequency) Average Rebuffering duration (pause duration) Pause intensity 17
How to Ensure QoEin Mobile Networks Are there any effective QoE metrics available? Server-1 enodeb? Server-n?...? Where to take QoE measurements? UE-1 (playback) UE-n (playback) 18
Recommendation ITU-T P.1201 (2012) Parametric assessment of audiovisual media streaming quality Objective Non-intrusive For UDP/IP Fidelity based 19
QoE Metric - Pause Intensity PI = 1 η/λ Determined by traffic property (η ) and service grade (λ). Represents the relative effectiveness of throughput η compared to the required playout rate λ. An objective and non-intrusive metric. For TCP/IP (HTTP/TCP/IP DASH) Continuity based T. Porter, and X.-H. Peng, An objective approach to measuring video playback quality in lossy networks using TCP, IEEE Communications Letters, Vol. 15 Issue 1, Jan. 2011, pp. 76-78. M. Seyedebrahimi, C. Bailey, and X.-H. Peng, Model and performance of a no-reference quality assessment metric for video streaming IEEE Trans. on Circuits and Systems for Video Technology, December 2013 20
Subjective Test: MOS vs. Pause Intensity PI is closely correlated with MOS 21
MOS vs. Pause Intensity (with different type of videos) The correlated of PI with MOS is content independent 22
Subjective Test: MOS vs Buffer Behaviour Metrics M O S 5 4 3 M R1 N C M O S 5 4 3 M R1 N C M O S 5 4 3 M R1 N C 2 2 2 1 0 0.2 0.4 0.6 (a) Pause Intensity (PI) 1 0 0.2 0.4 (b) Pause Frequency 1 0 5 10 15 (c) Pause Duration (sec) 23
QoE-Driven Resource Allocation & Optimisation Through machine learning Server-1 AI enodeb PI Server-n Low computational overhead in PI measurement for real-time QoE monitoring and control through mobile device, base station and backbone network. UE-1 (playback) UE-n (playback) 24
QoE Driven Resource Allocation Algorithm Aim to minimize the pause intensity of the users and maximize their service continuity.!" =arg # =$ %., R ' =. = R 25
PI-Controlled Scheduling Approach QoE evaluation on the: network side user side ()=*+, - =*+. /0 / 0 / 1 λ: video code rate; γ: transmission effect; η: throughput; R: total allocation; R : scheduler incoming rate 26
Performance of High-Order PI 27
Service Distributions with Control by PI Metrics 7 600 600 400 6 400 400 200 5 200 200 0 y 4 0 0-200 3-200 -200-400 2-400 -400-600 1-600 -600 0 x 500 600 0-500 0 x 500 PI and high-order PI metrics are used to achieve the desired QoE in terms of capacity, coverage, fairness and cost-effectiveness. 400 200 0-200 -400-600 -500 0 x 500-500 0 x 500 0 x 500 600 400 200 y -500 y 5 600 y y x 10 0-200 -400-600 -500
Content Delivery Scenarios
PSNR of Videos PSNR Loss free video Video with loss i i+m Frame number
PSNR Utility of Video Partition PSNR Loss free video Video with loss PSNR profit i i+m Frame number
Utility of Cached Video Partition PSNR i i+m s = i = ( PSNR PSNR ) Pf, p 1 i i,( f, p)
Optimal Caching To maximize the video utility of a cache server (network) based on the 0-1 knapsack problem. max s. t. n i= 1 n i= 1 [ P( w ( s i i, t i ) C ) f ( w i, v i )] P(.) -- video utility function f(.) -- distribution function of PI w weighting factor of cached content t partitioning factor v -- content popularity factor s size of cached content
Number of Partitions Stored
Video Utility for Client and Server Client Video utility Server Video utility
QoE Driven Rate Adaptation for Video Streaming (http-based adaptive rate streaming) Main Server Edge Server R i User R i < (1- PI th )*R i PI > PI th PI measurement can help determine when the rate needs to be changed and by how much as it takes account of the network statistics, required data rate and other users status. 36
QoE Driven Cache Management max s. t. To maximise the expected QoE by optimally caching video content files. The QoE utility model is determined by - Pause Intensity - playback rate allocated to users - Content popularity n i= 1 n i= 1 [ Q( PI [ βr i i, α, i u ( PI) + γ ] i ) f ( PI)] C Q(.) -- QoE utility function f(.) -- distribution function of PI R(.) -- playback rate of cached content α -- rate scaler β -- content popularity factor 37
Optimisation Strategy Requested rates r N = ( r, r2, L, r 1 N ) PI N = ( PI, PI 2, L 1 N, PI ) PI of cached video Popularity probability p N = ( p, p2, L, p 1 N ) R N = ( R, R 2, L 1 N, R ) Rates of cached video max s. t. n i= 1 n i= 1 [ Q( PI [ βr i i, α, i u ( PI) + γ ] i ) f ( PI)] C Considering trade-offs between: Performance and Fairness Continuity and Fidelity 38
Remarks and Summary Pause intensity is closely correlated with subjective results on perceived quality by viewers and this performance is content independent. QoE can be directly measured on both the network (enodeb) and end-user (UE) sides via PI. PI is capable of regulating the balance between fairness and efficiency in resource allocation, and ultimately increasing network capacity. PI can control the timing and amount of changes in bitrate for DASH. PI can aid the design and optimisation of cooperative cache networks for content distribution, prefetching and replacement, in conjunction with MDS coding and collaborative MIMO technologies. 39
Thank You Q&A 40