David Bowler
University College London
Density functional theory (DFT) has become an almost universal tool in areas as diverse as physics, chemistry, biochemistry and earth sciences. However, many calculations which involve DFT impose a significant approximation: the size of the system addressed is confined to a few hundred atoms, introducing an important and often uncharacterised source of error. Teams from NIMS and UCL have developed the CONQUEST DFT code specifically to address these issues. Using a recently-developed approach (multi-site support functions or MSSF), the code is capable of performing high accuracy calculations on up to 10,000 atoms, while with a linear scaling approach we have demonstrated calculations on more than 2,000,000 atoms. The code is also extremely efficient on high-performance computing, showing perfect weak scaling on nearly 200,000 cores of the K computer.
As we prepare to release the code under an open source licence, I will give an overview of issues that affect the computational cost of DFT calculations, and describe the ideas that underlie MSSF and linear scaling as well as their implementation in CONQUEST. I will also report on the performance of the code and give some examples of recent applications.