Automatic data distribution and load balancing with space-filling curves:: implementation in CONQUEST

We present an automatic, spatially local data distribution and load balancing scheme applicable to many-body problems running on parallel architectures. The particle distribution is based on spatial decomposition of the simulation cell. A one-dimensional Hilbert curve is mapped onto the three-dimensional real space cell, which reduces the dimensionality of the problem and provides a way to assign different spatially local parts of the cell to each processor. The scheme is independent of the number of processors. It can be used for both ordered and disordered structures and does not depend on the dimensionality or shape of the system. Details of implementation in the linear-scaling density functional code CONQUEST, as well as several case studies of systems of various complexity, containing up to 55 755 particles, are given.