The near sighted principle, that quantum interactions in insulating systems are local, has over the last decade guided the development of numerous "fast" numerical methods for ab initio quantum chemistry that achieve a reduced, O(N) cost with system size N. In addition to linear scaling, the near sighted principle also promises O(1) communication costs with respect to N/p and hence scalability if locality can be exploited in a careful way.
In this talk I will briefly overview the MondoSCF project for O(N) quantum chemistry and its current and future capabilities. I will then outline a global strategy for achieving scalable parallelism that extends current methods of irregular parallel computation. These involve ordering schemes based on a novel space filling curve to achieve data locality, new data structures to support early onset linear scaling and fine grained decompositions, and methods for distributed dynamic load balancing.
Copyright. All Rights Reserved.