for Modeling and Optimizing Distributed and Dynamic Multimedia - PowerPoint PPT Presentation

http:// medianetlab.ee.ucla.edu Towards a Systematic Approach for Modeling and Optimizing Distributed and Dynamic Multimedia Systems Presenter: Brian Foo Advisor: Mihaela van der Schaar

Proliferation of multimedia applications Video compression Multimedia stream mining Online gaming Postprocessing Image/video retrieval Virtual reality and 3D

Challenges for designing and optimizing multimedia systems • Multimedia data and applications are highly dynamic! – Real-time system resource adaptation required • Support for multiple concurrent applications – Dividing resources efficiently and fairly among applications – Regard for applications’ autonomy • Distributed computing resources – Collaboration required to jointly process application – Information-decentralization (delay, high communications cost, proprietary or legal restrictions)

Overview of Thesis Topics Proposed Challenges Applications Res. Mgmt. Framework RDC Modeling for RDC Modeling for Wavelet Coders Wavelet Coders Dynamic Dynamic Stochastic/analytic Stochastic/analytic Source/Workload Qualifying exam Source/Workload Models Models Characteristics Characteristics Model-based DVS Model-based DVS for Video Decoding for Video Decoding Modeled Parameters Decentralized Algs. Decentralized Algs. Private Application Private Application for Resource Mgmt. of for Resource Mgmt. of Utility Functions Utility Functions Decentralized Multiple Applications Multiple Applications Information Exchange Optimization (Tax Functions) (Information Exchange) Multiple Apps, Multiple Apps, Configuring Cascades Configuring Cascades Multiple Processors Multiple Processors of Classifiers of Classifiers Info. about Other Sites Safe Exp./Local Search Safe Exp./Local Search Focus of the talk for Distr. Classification for Distr. Classification Collaboration between Collaboration between Distributed and Multi-agent Learning Distributed and Multi-agent learning Autonomous Sites Autonomous Sites Rules-based Rules-based Decision Making Decision Making for Distr. Classification for Distr. Classification

Outline of Presentation • New area emerging: Resource-constrained stream mining – Static stream, same site – Static stream, different autonomous sites – Dynamic stream, different autonomous sites • Decentralized Resource Allocation for Multiple Multimedia Tasks – Tax functions • Modeling multimedia data and application dynamics – Applications to Dynamic Voltage Scaling • Conclusions and future directions

Cascaded Topologies of Classifiers on Distributed Stream Mining Systems: Same Site y Little League? Processing node 2 n Borealis, Aurora, TelegraphCQ y y Baseball? Basketball? Cricket? n In cooperation with Marvel and the y Team Sport? Winter Sport? Racquet Sport? System S Stream Processing Core group n n at IBM T. J. Watson, Hawthorne, NY Tennis? n [Foo, Turaga, Verscheure, vdSchaar, y Ice Sport? Amini, Signal Processing Letters, 2008.] n Processing node 3 Skiing? y n Skating? y Processing node 4 • Complex classifiers can be decomposed into cascaded topologies of binary classifiers [Schapire, 1999] • Application operators can be instantiated on distributed processing devices with individual resource constraints. • Issues: placement, fault tolerance, load shedding, etc.

Prior Approaches to Load Shedding for Stream Mining Systems • Probabilistic load shedding – Reduce network delay [Tatbul 2002] – Reduce memory consumption [Babcock 2003] • Quality-aware load shedding for data mining – Windowed load shedding for aggregation queries [Tatbul 2004, 2006] • Load shedding for classification? Very little work in this area! [Muntz 2005] – Single classifier – • Limitations – Suboptimal classification performance/application utility! • Our approach – First to formalize load shedding as an application optimization problem: maximize joint classification quality subject to resource constraints, delay, and dynamics

Configuring Classifier Operating Points False alarms False alarms Miss [pos class] Miss [neg class] SVM: Linear Kernel Function SVM: Radial Basis Kernel Function e   u  v 2 T     u v k k k k P th th D 1 DET curve relates misses and false alarms. Can parameterize operating point by pf. Affects throughout (output rate) P 0 F and goodput (output rate of detected data) 0 1

Problem Formulation • Given – Costs of misclassification ( c M , c F ) per data object per class k  – True volume of data in each class – Placement and Resource constraints – Throughput and Goodput • Objective – Minimize end-to-end misclassification cost – Satisfy resource constraints R 2 1 1 1  c , c (Throughput, Goodput) F M K t g ,   y 2 2   k k  minimize c  c  false_alarms  c  misses   F M n R 1 k  1   t g ,  t g ,  2 2 2  c , c 2 2 1 1 K F M y       k k k k k k   c  t p  g p  c    g p         F F F M F k  1 n 3 3 3  c , c F M    t g ,  t g ,  s.t. Ah p R   3 3 y 1 1 F   t g , 0  p  1 n 3 3 F 4 4 4  c , c F M

Computing Goodput and Throughput for a Semantic Tree Topology y Little How can goodput/throughput be computed? League? n y y Baseball? Basketball? Cricket? n y Team Sport? Winter Sport? Racquet Sport? n n Tennis? n y Ice Sport? n Skiing? y n Skating? y When each classifier in a branch filters a subclass, this property is referred to as exclusivity. Goodput and throughput for each class can be computed recursively.

Calculation of throughput and goodput t   i   C   g C j  i  i   ˆ i i     X   Pr{ X } i Pr p i 0 0 ˆ D i t X     anc i i 0   i   p X F i   g  anc i      ˆ i X   i p Pr ˆ i 1   X   F 0 i i 0 i p D   t C i   j   g    i  Classifier-describing matrices:       i i i     i i i p p p   p p p D i F D   F i D F   k or k     T T   i i i i   0 p  0 p       i D  i D Throughput and goodput out of classifier Ci   1 t t     i anc i         k k k k    T ... T T ... T   i i 1     g g anc( ) i 1    i   anc i     

Calculation of resource constraints h C • Resource consumed by classifier is i i t proportional to the input rate , i.e. anc i     h t i i anc i    – Coefficient is the processing complexity per data i object • Placement: described by matrix A, where:   A 1 if C node( ) m mi i  A 0 otherwise mi T  R R R  1 ,...,  • Node resource availability: M  Ah R • Resource constraint inequality:

Prior Approaches to Load Shedding for Stream Mining Systems R 1 To reduce delay when not be feasible to meet tight Operating Operating resource constraints  Point 1 Point 2 Arbitrary Load Shedding at next classifier [Babcock, 2003; Tatbul, Zdonik, 2006] R 2 Operating Region 1 DET curve 0.9 0.8 0.7 0.6 Random data forwarding curve p D 0.5 0.4 0.3 0.2 0.1 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 p F

A Proposed Approach: Multiple Operating Points Shedding of low confidence data at current classifier  Intelligent Load Shedding [pos class] Also support Replication of low confidence data when resources are available [neg class] (significant difference from current literature) Positive threshold Negative threshold

Centralized Solutions • Use Sequential Quadratic Programming (SQP). – Running SQP several times with different starting points gives higher probability of finding global optimum. • Considered Algorithms – A) Equal Error Rate (EER) configuration (e.g. [Muntz, 2005]) – B) Single operating point, no resource constraint consideration. • Let system take care of load shedding • (e.g. [Schapire, 1999] + [Babcock, 2003; Zdonik, 2006]). – C) Single operating point, jointly optimized by shedding load at the output [Foo, Turaga, Verscheure, vdSchaar, Amini, SPL, 2008.] • Algorithm considers resource constraints downstream and configures operating point and load shedding jointly. – D) Proposed: Multiple operating points! [Foo, Turaga, Verscheure, vdSchaar, Amini, SPL, 2008.] • Use a separate threshold for yes and no output edges. • Intelligent load shedding and replication of data! Distributed algorithm? [Foo, Turaga, Verscheure, vdSchaar, Amini, TCSVT, submitted 2008]