14
Parallelism Real Time Parallelism Real Time Anca Batori Lavinia Basarabă Anca Brandimbur “Politehnica” University of Timişoara Automation and Computing University Computers Department iunie 2010

Parallelism Real Time

Embed Size (px)

DESCRIPTION

“Politehnica” University of Timişoara Automation and Computing University Computers Department. Parallelism Real Time. Anca Batori Lavinia Basarabă Anca Brandimbur. iunie 2010. Objectives. - PowerPoint PPT Presentation

Citation preview

Page 1: Parallelism Real Time

Parallelism Real TimeParallelism Real Time

Anca BatoriLavinia

BasarabăAnca

Brandimbur

“Politehnica” University of TimişoaraAutomation and Computing University

Computers Department

iunie 2010

Page 2: Parallelism Real Time

ObjectivesObjectives

presenting the use of graphical processing units (GPU) to achieve significant improvements for real time systems

presenting the main steps for developing an application using CUDA

offering a source of resources

Page 3: Parallelism Real Time

CConontentstentsIntroductionParallelism Application descriptionExperimentsConclusions

Page 4: Parallelism Real Time

IntroductionIntroductionReal Time Systems are an

important area of research and development

Many applications: Airplanes, biomedical accelerators, nuclear power plants

Necessity of parallelism to achieve desired time limits

Page 5: Parallelism Real Time

ParallelismParallelismCan be hardware and software

◦GPUs represent a combinationGPUs have certain characteristcs,

that CPUs do not poses, that can be useful for certain application◦More processing power, less

flexibilityThe application are developed

using CUBLAS

Page 6: Parallelism Real Time

AArrcchitectuhitecturesresCPU CPU vsvs GPU GPU

Cores number 4Threads 2Cache memory

=> random address acces

Cores number 240Threads 1024Cache memory =>

fast access to consecutive addresses

Sursa: www.nvidia.com

Page 7: Parallelism Real Time

Support Vector Machines Support Vector Machines (SVM)(SVM)

Possibilities to split two classes

Page 8: Parallelism Real Time

SVMSVM geometrical geometrical representationrepresentation

Page 9: Parallelism Real Time

CPU versus GPUCPU versus GPUSpeedup 54x for1000 imagesEven better results for bigger training

sets

Page 10: Parallelism Real Time

Multiclass Multiclass SVMSVM

“One Against All” (OAA)training: M binary classifiers(M number of

classes)testing: strategy„the winner takes it all”

“One Against One” (OAO)training : M(M-1)/2 binary classifiers testing : strategy maximum number of votes

“Directed Acyclic Graph” (DAG)testing : decision tree

Page 11: Parallelism Real Time

Multiclass AlgorithmsMulticlass Algorithms

NVIDIA GTX 280IDSIA sets

Training set dimension

Training time (secunde) Recognition percentage (%)

OAO DAG OAA OAO DAG OAA

100 15,183 15,183 7,092 84,94 84,45 67,39

300 129,188 129,188 77,438 88,37 87,91 82,19

1000 1806 1806 1452,7 90,73 89,84 82,94

Page 12: Parallelism Real Time

Multiclass Multiclass AlgoritAlgorithhmmss

Page 13: Parallelism Real Time

ConclusionsConclusionsIn the period of ubiquitous and

pervasive systems, real time systems are a very important field

Since CPUs do not evolve, so fast as they used to, there is a shift towards parallel programming and more and more systems are developed this way

GPUs can improve the execution time for an application 100x times

Page 14: Parallelism Real Time

Thank youThank you