Continuing advances in computer hardware enable yet bigger systems to be simulated using high-level ab initio techniques such as coupled-cluster theory which represents the �gold standard� in quantum chemistry. However, during the past several years increases in single-core computing performance have stalled due to physical limits on transistor sizes and clock speeds achievable at current manufacturing technology. Even on mobile devices the dominant increase in computational power comes not due to improved single-core performance but because of increasing number of processing units. Growing interest in energy efficient computing also contributes to the trend of using many slower but less energy hungry processors at supercomputing facilities in recent years. The state-of-the-art supercomputers have hundreds of thousands of processing cores with projections predicting millions of cores at exascale computing facilities within next five years. Moreover, to achieve even more efficient utilization of computer resources most computationally intensive tasks are offloaded to more energy efficient specialized hardware accelerators. This presents a challenge of adapting high-level quantum chemistry techniques to modern multi-node/multi-core/accelerator-enabled architectures. The fact that these methods require manipulations of huge multi-dimensional tensors poses another challenge in designing algorithms to minimize costly data transfers between the nodes and between the accelerator and main memory within a node. Here we present our work on parallelization of libtensor � the tensor library used in state-of-the-art Q-Chem quantum chemistry package. We discuss the strategies to minimize data transfer and enable yet bigger ab initio calculations on future multi-node/multi-core/accelerator-enabled platforms as well as the need for fault-tolerant approaches to mitigate inevitable hardware failures on such scale.

Parallelization of tensor software for high-level <i>ab initio</i> calculations on future computing platforms

248th National Meeting (2014)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.

Parallelization of tensor software for high-level ab initio calculations on future computing platforms