Algorithms and Software for Quantum Chemistry at Petascale and Beyond
Quantum chemistry faces many challenges on petascale supercomputers due to a changing landscape of processing architecture and the enormous scale of these systems. Over the past 10 years, systems have grown in size by a factor of a thousand in some cases and within a node there is significantly more parallelism. This talk will start by considering the NWChem software package and it's coupled-cluster (CC) codes, which are scalable to 100,000 cores on Cray supercomputers and have rich scientific capability.
From here, we consider three research projects to develop new coupled-cluster algorithms for modern supercomputers. The first of these is the development of CC for GPUs, where we observe approximately an order-of-magnitude speedup with respect to existing codes for CCSD at workstation scale. The second project is a complete redesign of tensor contraction algorithms using novel techniques from dense linear algebra. This completely eliminates load-imbalance and allows for topology-aware mapping on systems with torus interconnects (e.g. Blue Gene and "K"). Finally, we demonstrate how inspector-executor techniques can be used to eliminate dynamic task scheduling in NWChem CC codes, which can lead to a substantial reduction in time-to-solution.
This work has been done in collaboration with Eugene DePrince (Georgia Tech), Edgar Solomonik (Berkeley), Devin Matthews (Texas), David Ozog (Oregon), Pavan Balaji (Argonne) and Jim Dinan (Argonne).