Software Libraries for Message Passing Computers

David W. Walker

1 Title page

2 Overview of talk

3 Motivation for parallel libraries

4 Desirable features of parallel libraries

4.1 Communication safety

4.1.1 Intended behaviour of example 1

4.1.2 Unintended behaviour of example 1

4.1.3 Intended behaviour of example 2

4.1.4 Unintended behaviour of example 2

4.2 Process groups

4.3 Application topologies

5 Some existing parallel libraries

5.1 The Parallel Utilities Library

5.2 The Multicomputer Toolbox

5.3 The PETSc toolkit

6 The ScaLAPACK parallel library

6.1 Problems addressed by ScaLAPACK

6.2 Key ideas in ScaLAPACK

6.3 Non-uniform memory access machines

6.4 Unpartitioned LU factorisation

6.5 Block partitioned LU factorisation

6.6 The block-cyclic data distribution

6.7 Example of a block-cyclic data distribution

6.8 ScaLAPACK building blocks

6.8.1 Black box and tool box use modes

6.8.2 Basic Linear Algebra Communication Subprograms

6.8.3 Parallel BLAS

6.8.4 PBLAS operations

7 Approaches to parallel I/O

7.1 Basic parallel I/O operations

7.2 I/O cost of out-of-core LU

7.3 Hybrid LU factorisation algorithm

8 Object-oriented approaches

9 Procedural and object-oriented approaches

10 Summary and future work