|
Efficient numerical algorithms lie at the heart of every successful
scientific application program. The CSE Department has a long track record in algorithm
development tailored to specific tasks in science and engineering.
The Advanced Research Computing Group is developing advanced algorithms
to enhance the performance of other programmes at Daresbury.
A parallel 3D FFT package has been written which addresses directly problems
found using other FFTs in the smooth particle mesh Ewald (SPME) algorithm
used within DL_POLY. This allows high performance to be obtained without extensive
data redistribution.
The Block Factor Jacobi (BFG) package implements a version of the Jacobi algorithm
for matrix diagonalization, which offers, compared to previous code,
extended functionality, better numerical stability, increased single processor performance
and better scaling. It can diagonalize both real symmetric and Hermitian matrices,
and has interfaces for matrices distributed either in a block-cyclic fashion,
or by columns. It is now being used within CRYSTAL and has been distributed
to the authors of Siesta.
|
