Next:
Contents
Contents
Bands-parallelism in Castep
A dCSE Project
Phil Hasnip
Contents
Introduction
The dCSE Project
Summary of Progress
Castep
Parallelisation
Computational Costs
Orthogonalisation and Diagonalisation
Fourier Transforms
Castep Performance on HECToR
(Work Package 0)
General Castep Performance
Benchmarks
FFT
Maths Libraries (BLAS)
Compiler
Node Usage
Baseline
Analysis
Band-Parallelism
(Work Package 1)
Introduction
Programming
Comms
Ion
Optimisation
Wave
Band-overlaps and dot-product
Rotations
Communication pattern
Optimisation Note
wave_(S)dot_lower
Type issues
wave_write
Band distribution
Known issues remaining
Density
Ewald
Electronic
Model
Secondd
Phonon
Testing
Benchmarking and Performance
Scope of Parallelism and Integration with main Castep Codebase
Distributed Diagonaliser and Inverter
(Work Package 2)
Introduction
Programming
Development
ScaLAPACK Performance
Castep Implementation
Castep Performance
Limitations
Independent Band Optimisation
(Work Package 3)
Introduction
Castep Eigensolver
Proposed Eigensolver
Performance
Other Developments and Future Work
-Point
Distributing
-Projectors and
Mpitrace
Final Thoughts
This Project
HECToR as a Development Machine
Bibliography
About this document ...
Sarfraz A Nadeem 2008-09-01