T chnology chnology Ma turity turity for fo Adaptiv Adaptiv Massively Massiv ely Pa P ra r llel llel Computing F rst rst Wo W rksho p 2009

Transcription

1 Technology Maturity for Adaptive Massively Parallel Computing First Workshop 2009 March h2 3, 2009 Portland, OR, USA

2 AMP Computing Workshop 2009 Massive Data Computing Pradeep K. Dubey Senior Principal Engineer Intel Corporation 2

3 Norman s Gulf Execution Gap Computer s System Model Human s Conceptual Model Evaluation Gap

4 What is Is it What if? Large dataset mining Semantic Web/Grid Mining Streaming Data Mining Mining Distributed Data Mining Content-based Retrieval Multimodal event/object Recognition IndexingCollaborative Filters Statistical Computing Multidimensional Indexing Machine Learning Streaming Dimensionality Reduction Clustering / Classification Dynamic Ontologies Model-based: Efficient access to large, unstructured, sparse datasets Recognition Bayesian network/markov Model Stream Processing Neural network / Probability networks Photo-real Synthesis LP/IP/QP/Stochastic Optimization Real-world animation Synthesis Ray tracing Global Illumination Behavioral Synthesis Graphics Physical simulation Kinematics Emotion synthesis Audio synthesis 4 Video/Image synthesis Document synthesis

5 Interactive RMS Recognition Mining Synthesis What is? Is it? What if? Model Find an existing model instance Create a new model instance Most RMS apps are about enabling interactive (real-time) RMS Loop or irms 5

6 Visual Computing Loop Recognition Mining Synthesis What is? Is it? What if? Model Find an existing Create a new model instance model instance Graphics Rendering + Physical Simulation Learning & Modeling Computer Vision Visual Input Streams Reality Augmentation Synthesized Visuals

7 Analytics Loop Recognition Mining Synthesis What is? Is it? What if? Model Find an existing Create a new model instance model instance Semantic Web Mining Learning & Filling Ontologies Mining Structured Data + Unstructured Blogs Structured Augmentation Synthesized Structures

8 RMS Inner Loop Unrolled Procedural or Analytical Visual Computing (Graphics and Vision) Model Physical Simulation Rendering Data Acquisition and Mining Based Non-Visual Computing (Search and Analytics) Procedural or Analytical l Model Real-time Indexing What-if evaluation Summary synthesis Data Mining Based (Crawling)

9 Visual Computing Meets Analytics Rendering Simulation Machine learning Neural networks Probabilistic reasoning collision detection force solver global illumination Physics Fuzzy logic Belief networks Evolutionary computing Chaos theory Soft Computing Soft Physics? Dynamics Constraints Constraint Dynamics

10 Nested irms Recognition Mining Synthesis What is? Is it? What if? Graphics Rendering + Physical Simulation Semantic Web Mining Learning & Filling Ontologies Mining Structured Data + Unstructured Blogs Structured Augmentation Synthesized Structures Learning & Modeling Computer Vision Visual Input Streams Reality Augmentation Synthesized Visuals

11 Connected Computing Visual Computing (Ocean of Clients/Devices) Graphics Rendering + Physical Simulation private data, sensory inputs, streams/feeds immersive 3D graphics Semantic output, Web Mining interactive visualization Learning & Filling Ontologies Structured Data + Unstructured Blogs Structured Augmentation Massive Data Analytics Mining (Cloud of Servers) Synthesized Structures Learning & Modeling Intersection of massive data with massive compute real-time analytics, massive data mining-learning Computer Vision Visual Input Streams Reality Augmentation Synthesized Visuals Architectural implications are far more radical Computational substrate must undergo a sea-change! 11

12 Immersive Computing Sensory Immersion Behavioral Immersion Rendering Simulation Super Immersion collision detection force solver global illumination Physics Machine learning Neural networks Probabilistic reasoning Fuzzy logic Belief networks Evolutionary computing Chaos theory Trade Bots Soft Computing Outer loop: irms Visual Soft Loop: One Per Physics? Real User Dynamics Chat Bots Inner Constraints Loop: irms Analytics Loop One per bot per user Performance Needs: Can far exceed typical i/o limits of human perception Shop Bots Constraint Dynamics Computational requirements are huge, but Gambler Bots C t ti l i tplayer Bots h b t Pradeep K. Dubey Reporter Bots pradeep.dubey@intel.com Performance Needs: Limited by input/output limits of human perception 12

13 Data changes the game 1997 Today [Kasparov vs. Deep Blue] Rule-based system exceeds human performance in a structured, deterministic domain [Google MT wins NIST contest] Statistical inference (not rules) 100s of TB of training data Racks of computation Newcomer Google beats decades of rule-based translation research Opportunities Abound: Massive Data with Massive Compute 13

14 Machine Learning Algorithm Classes Model based Transparent model Bayes nets Regression Discriminant Analysis Opaque model Neural networks SVM Model free Supervised Classifiers Decision trees Ensembles of trees Unsupervised Clustering Association Rules Sequential Patterns Principal Components K Nearest Neighbor 14

15 What are we doing? Rendering Video Miningi Portfolio RT/Global Illum. Physical Simulation (Patterns) Surv. Camera (Anomaly) Selection Asset Allocation Behavioral Net Intrusion Derivative Simulation Detection Pricing Offline/Interactive Trading Floor Asset-Liability Cinema Data Streams Management Data Tracking and Credit-Card Visualization Trading Fraud Detect Visual Risk Web Trans Computing Management Streams Multi-Party Interest-Rate Sensor Data Games Models Streams Virtual Worlds Multi-party Medical Data Simulation i Auctions Ad-hoc Search Named Entity Extraction Automatic Structuring Machine Translation Collaborative Filtering Fact-Event Mining Multimodal Search Semantic Search Personal Streams Agent / Spider Computer-Aided Model Diagnostics / Auto Bots Virus/Spam Calibration Treatment Workload-driven Architecture Research Unstructured Graphics, gaming and Financial Analytics Graphics Rendering Physical Stream Mining Simulation Real-time -- Vision decision-making i Data Mining -- Information Analytics SAAR (Scalable Applications and Architecture Research) Management Multi-party t collaboration 15

16 Scaling With Cores Par rallel Spe eedup 70 Production Fluid Production Face 60 Production Cloth Marching Cubes 50 Sports Video Analysis Video Cast Indexing Text Indexing 40 Foreground deti Estimation Human Body Tracking 30 Portifolio Mangement 3D FFT 20 BLAS Number of Cores

17 Architectural Smarts

18 RMS Computing Core: Algorithmic Evolution Map based shading Global Illumination based SIMPLEX based linear optimization Mass Spring based ddf deformation Marker based explicit surface tracking Linear manifold based recognition/modeling Linear Complementarity problems Low dimension classifiers IPM based LP/QP/NLP optimization FD/FE/FV based deformation Non-Linear Level Set And based implicit surface Generative tracking Non linear manifolds computer vision Non linear Complementarity Low dimension classifiers High dimension classifiers

19 Summary Connected Computing It s all about three C s (above + content or data) Architectural lchallenge Moving the data real time to where compute happens Algorithmic Opportunity Massive data approach to traditional compute problems

20 AMP Computing Workshop 2009 BACKUP 20