CRD Researchers Nominated for Best Paper at ICPP2019

Researchers in Lawrence Berkeley National Laboratory’s Computational Research Division are developing new Eigensolver algorithms for electronic structure codes based on an iterative approach using an unconstrained energy functional scheme to improve scaling on parallel computing architectures. Eigensolvers are important in many scientific codes used in materials, chemistry, and structural and data analysis. 

This research, by CRD’s Mauro Del Ben, Osni Marques, and Andrew Canning, is being recognized with a best paper nomination at the International Conference on Parallel Processing (ICPP2019) taking place this week in Kyoto, Japan.

“As we move toward the exascale era of computing, we need Eigensolvers with highly efficient parallel scaling to exploit these types of new, highly parallel architectures,” Canning said. “Compared to currently used constrained and direct methods, the unconstrained approach offers much better scaling properties for solving large problems on parallel computers.”

The CRD researchers implemented their formulation in the first-principles materials and chemistry codes CP2K and Quantum Espresso. Their paper describing this method - which includes studies of its convergence and parallel scaling on Cori, the Cray XC40 supercomputer at NERSC - was selected as the best paper candidate at ICPP2019 in the applications track out of 75 papers submitted, with 22 papers accepted. Marques is presenting the paper at the conference this week.

Water system used for parallel scaling tests on up to 230k cores on the Cori Cray XC40 system at NERSC.

“ICPP is a difficult conference to get a paper accepted, and being a candidate for best paper is an even bigger accomplishment,” said Horst Simon, deputy director for research at Berkeley Lab.

This project was initially funded by the Laboratory Directory Research and Development Program and has applications for SciDAC projects where these types of codes, such as CP2K, are used.