COSC 6374 Parallel Computation. Organizational issues (I)

Transcription

1 COSC 6374 Parallel Computation Spring 2007 Organizational issues (I) Classes: Monday, 4.00pm 5.30pm, SEC 204 Wednesday, 4.00pm 5.30pm, SEC 204 Evaluation 2 homeworks, 25% each 2 quizzes, 25% each In case of questions: gabriel@cs.uh.edu Tel: (713) Office hours: Mo, Wed, 2pm-3.00pm or by appointment All slides available on the website: TA: Hatem Ltaief (ltaief@cs.uh.edu) 1

2 Why Parallel Computing? To solve larger problems many applications need significantly more memory than a regular PC can provide/handle To solve problems faster despite of many advances in computer hardware technology, many applications are running slower and slower e.g. databases having to handle more and more data e.g. large simulations working on even more accurate solutions Parallel Programming Exploit concurrency Internet: Client and server are independent, interacting applications Searching an element: distribute the search database onto multiple processors Adding two arrays of integers: Processor 1 + = + = Processor 2 + = 2

3 Scalar product: Parallel Programming (II) s = N 1 i= 0 Parallel algorithm s = N / 2 1 i= 0 N / 2 1 a[ i]* b[ i] ( a[ i]* b[ i]) + N 1 i= N / 2 = ( alocal[ i]* blocal[ i]) i= rank = 0 ( a[ i]* b[ i]) + N / 2 1 ( alocal[ i]* blocal[ i]) i= rank = 1 requires communication between the processes Flow around a re-entry vehicle 3

4 Simulation of combustion processes Simulations of airplanes 4

5 Determining phylogenetic trees HPC in film industry Special effects are highly compute intensive Example: Lord of the Rings company: Weta Digitals 3200 processor cluster a single scene contains: per second 24 frames per frame: 4996 x 3112 points with 32- or 64 bit color encoding each object means a separate compute cycle Number of computer-added special effects in movies: Jurassic Park : 75 Lord of the Rings (I): 540 each of the following episodes of Lord of the Rings doubled the number of special effects last episode of Star-Wars:

6 HPC in financial companies large data-bases simulations for risk assessment of portfolios each access (=web search) means that several hundreds MB of data have to be touched Google s database consists of hundreds of terabytes of data A web search is trivially parallel Google s cluster-philosophy: many cheap (unreliable) nodes Reliability achieved through replication in software and hardware End of 2003: several PC cluster with each nodes 6

7 Luis Andre Barrosso, Jeffrey Dean, Urs Hölzle, Web search for a planet: The Google Cluster Architecture, IEEE Computer Society Multi-core processors 7

8 Parallel Computer Parallel Computer PC cluster in U height per node - Myrinet Network - Fast Ethernet Network (1U = 1.75 = 4.45 cm) 8

9 Parallel Computer PC Cluster in U height - Infiniband Network - Gigabit Ethernet Network 9

10 Earth Simulator Target: Achievement of high-speed numerical simulations with processing speed of 1000 times higher than that of the most frequently used supercomputers in ( 640 nodes 8 processors/node 40 TFLOPS peak 10 TByte memory Top 500 List ( 10

11 Top 500 List Accompanying literature (I) Ananth Grama, Anshul Gupta, George Karypis, Vipin Kumar: Introduction to Parallel Computing, Pearson Education, John May: Parallel I/O for High Performance Computing, Morgan Kaufmann, H. Charles Romesburg, Cluster Analysis for Researchers, LuLu Press North Carolina,

12 Accompanying literature (II) Accompanying literature Timothy G. Mattson, Beverly A. Sanders, Berna L. Massingill Patterns for Parallel Programming, Software Pattern Series, Addison Wessley, Jack Dongarra, Ian Foster, Geoffrey Fox, William Gropp, Ken Kennedy, Linda Torczon, Andy White Sourcebook of Parallel Computing, Morgan Kaufmann Publishers, Michael J. Quinn: Parallel Programming in C with MPI and OpenMP, McGrawHill, L. Ridgeway Scott, Terry Clark, Babak Bagheri: Scientific Parallel Computing, Princeton University Press, Outline of the course 1. Introduction to Parallel Computation 2. Parallel Computer Architecture 3. Parallel Programming MPI Algorithms Application domains and examples Classification of parallel algorithms Parallel I/O 4. Parallel Performance Analysis Performance modeling Introduction to the theory of cluster analysis 12