Linear Algebra Libraries for High-Performance Computers
PPT Slide
Heterogeneous Network Computing
History: Late 60’s - Early 70’s Wilkinson-Reinsch Handbook
History: Early 70’s NATS Project
History: Mid to Late 70’s EISPACK and LINPACK
High-Performance Computing Directions
Uniprocessor Cost: Reality
Memory Hierarchy
Performance = Effective Use of Memory Hierarchy
How To Get Performance From Commodity Processors?
Adaptive Approach for Level 3 BLAS
Code Generation Strategy
ATLAS Across Various Architectures 500x500 DGEMM
500 x 500 Double Precision RB Right-Looking LU factorization
Multithreaded BLAS for Performance
LAPACK Linear Algebra Library in F77
ScaLAPACK
Possible Data Layouts
Distribution and Storage
ScaLAPACK - What’s Included
ScaLAPACK Structure
Parallelism in ScaLAPACK
Nonsymmetric Eigenvalue Problem
Parallelation
Performance of PDLAHQR
Symmetric Eigenvalue Problem
Basic Idea
Parallelization
Execution Times IBM SP2 - 8 Procs
Scalability
Additional Parts
Heterogeneous Computing Issues
ScaLAPACK Summary
Templates Project
Problem Solving Environments & Computational Grid
Network Enabled Servers
Motivation for NetSolve
NetSolve - MATLAB Interface
Hiding the Parallel Processing
Next Step: Hardware & Software Servers
Participants
References
Email: dongarra@cs.utk.edu
Home Page: http://www.netlib.org/utk/people/JackDongarra/