Overview: Why CONQUEST?¶
There are already many DFT codes which are available under open-source licences. Here we give reasons why you might choose to use CONQUEST.
CONQUEST is designed to scale to large systems, either using exact diagonalisation (with the multisite support function approach, we have demonstrated calculations on over 3,000 atoms) or with linear scaling (where calculations on over 2,000,000 atoms have been demonstrated). Moreover, the same code and basis sets can be used to model systems from 1 atom to more than 1,000,000 atoms.
CONQUEST is an inherently parallel code, with scaling to more than 800 cores demonstrated for exact diagonalisation, and nearly 200,000 cores with linear scaling. This scaling enables efficient use of HPC facilities. CONQUEST (in linear scaling mode, as well as to a certain extent for exact diagonalisation) scales best with weak scaling: fixing the number of atoms per core (or thread) and choosing a number of cores based on the number of atoms.
CONQUEST also offers some OpenMP parallelisation in linear scaling mode, with relatively low numbers of MPI threads per node, and further parallelisation performed with OpenMP.
The ideas of linear scaling have been current for more than twenty years, but it has proven challenging to make efficient, accurate codes to implement these ideas. CONQUEST has demonstrated effective linear scaling (with excellent parallel scaling), though is still somewhat restricted in the basis sets that can be used. For calculations beyond 5,000-10,000 atoms with DFT, linear scaling is the only option.
CONQUEST expresses the Kohn-Sham eigenstates or the density matrix (which are equivalent) in terms of local orbitals called support functions. These support functions are made from one of two basis sets: pseudo-atomic orbitals (PAOs) or blip functions (B-splines); the main basis functions in use in CONQUEST are the PAOs. A PAO generation code is included with the CONQUEST distribution, with well-defined and reliable default basis sets for most elements.
The simplest choice is to use one PAO for each support function (typically this allows calculations up to 1,000 atoms). For diagonalisation beyond this system size, a composite basis is used, where PAOs from several are combined into a smaller set of support functions (multi-site support functions, or MSSF). With MSSF, calculations on 3,000+ atoms are possible on HPC platforms. For linear scaling, more care is required with basis sets (more details can be found here).