Application-driven Cross-layer Optimization in Wireless Networks

Transcription

1 Application-driven Cross-layer Optimization in Wireless Networks Srisakul Thakolsri *, Wolfgang Kellerer * Shoaib Khan, Eckehard Steinbach * Future Networking Lab Ubiquitous Services Platform group DoCoMo Euro-Labs Media Technology Group Institute of Communication Networks TU München 2 nd Seminar on Service Quality Evaluation in Wireless Networks University of Stuttgart June 12 th, 2007

2 Outline Motivations Cross-Layer Optimization (CLO) Architecture Multi-application CLO Voice Streaming File transfer Simulation results

3 Motivations Video Streaming Server BS doesn t have knowledge of application layer File Sharing Server Mobile User Equipments Voice Communication End-devices Base Station + CLO Function Multiple users sharing wireless medium, e.g. in a cell, usually run different applications simultaneously Impact of losses on user-perceived quality is application-dependent Optimizing the system for different users and applications requires: 1. defining a common metric that quantifies the user satisfaction 2. mapping network and application parameters onto this metric.

4 Proposed CLO Architecture Mean Opinion Score Voice FTP Video Parameter Abstraction Cross-layer Optimizer ~ X PEP Mean Opinion Score Random packet loss, %Data rate, kbps ~ ~ = A R x~opt Decision PEP Parameter Abstraction ( transmission rate, packet error prob., packet size ) Mean Opinion Score A ~ Decision Distribution Decision Distribution R ~ Base Station Application Layer (Streaming Video) Transport Layer Network Layer Radio Link Layer Data Link Layer Physical Layer Kellerer, Choi, Steinbach, Khan WPMC'03, ICIP'04, IEEE ComMag'06

5 Cross-layer Optimizer: Maximization of User Satisfaction Common Metric? Voice FTP Video... U 1( R1, PEP1) U 2( R2, PEP2) U 3( R3, PEP3) Resource allocation ( R, PEP) ~ x = arg max U ( ~ x) opt ~ ~ x X i ~ X : set of operating modes U : utility function i / resource allocations Khan, Duhovnikov, Steinbach, Sgroi, Kellerer, ISMW '06

6 Challenge: estimating utility functions video streaming?? video conferencing?? voice Mean Opinion Score FTP(image download)

7 Utility Function (PESQ) of Voice Codecs Mean Opinion Score (MOS) G B (6.4 kbit/s) ilbc (15.2 kbit/s) SPEEX (24.6 kbit/s) G.711 (64 kbit/s) 0.35 MOS packet error prob. (PEP), %

8 Video streaming: Estimating reconstruction quality at the receiver Source Distortion due to compression Loss Distortion due to transmission loss DS Source Distortion D L Loss Distortion S Source Encoder Channel Encoder Wireless Channel Channel Decoder Source Decoder Rate Total Distortion D D s = D L = f f (R) ( R, PEP) D = D S + D L Packet Error Probability

9 Utility function of streaming video D = D S + D L PSNR = 10 log D MOS = a PSNR + b PSNR 40 db = MOS 4.5 PSNR 25 db = MOS 1 2 Mean Opinion Score Foreman H Packet Error Probability (PEP), %

10 Utility function of FTP MOS( R, PEP) = a log10 1 ( b R( PEP) ) A. Saliba et al Mean Opinion Score Packet error prob., % Data rate, kbps : user subscribed to R gets R 1: minumum R a, b Minimum (min MOS) Contract (max MOS)

11 Performance of Utility-based optimization Maximize mean MOS: 1 arg max ~ x X K K ~ k = 1 MOS ( R ( ~ x), PEP Maximize Throughput: arg max ~ ~ x X K k = 1 R k k ( ~ x) k ( 1 PEP ( ~ x) ) k k ( ~ x)) CDF MOS, 500Ksym/s MOS, 1000Ksym/s MOS, 1500Ksym/s Throughput, 500Ksym/s Throughput, 1000Ksym/s Throughput, 1500Ksym/s mean MOS Seven users: 3 voice users, 2 FTP users, 2 video users Total system rates 500, 1000, and 1500 ksymbols/sec Session duration: 30 sec Resource allocation update: every 1 second

12 Greedy Allocation Algorithm Greedy Optimization ΔU i max, { K} ΔU (, i j) 1,..., ΔU i : utility change due to an increase of ΔU j : utility change due to a decrease of Δα :change of resource i share for user i j i j Δα Δα i j mean MOS Total symbol rate = 1Msym/s greedy algorithm full search number of user 4 video users, 2 FTP, and (K-6) voice users Full search becomes computationally infeasible for K>10 Real-time optimization for greedy algorithm

13 Optimizing the number of users in a cell Here, we maximize the minimum MOS of the users. A target min. MOS is set at the beginning of the simulation Total symbol rate is fixed: 200Ksymbol/sec minmos maxminmos MaxThroughput Only voice service is considered number of user

14 Summary Using MOS as a unifying optimization metric across different types of application Defined techniques for mapping network and application parameters onto MOS Results show advantages of MOS-based approach comparing over throughput maximization approach Improve user-perceived quality Optimize usage of network resource

15 Thanks for your attention... Questions?