Murat Keçeli
Postdoctoral Appointee Murat Keçeli (CSE) will discuss his Laboratory-Directed Research and Development (LDRD) sponsored work at the LDRD Seminar Series presentation Tuesday, July 18, 2017.
“Toward Exascale Quantum Chemistry with Sparse Eigensolvers” begins at 12:30 p.m. in the Bldg. 203 Auditorium. All are welcome to attend.
Abstract
Widely used electronic structure codes are not able to use efficiently more than one percent of the cores available on leadership-class supercomputers. These codes are based on Hartree-Fock molecular orbital and/or Kohn-Sham density functional theories, which requires the repeated solution of an eigenvalue problem through many iterations until a self-consistent field (SCF) convergence criterion is met. Conventionally, dense linear algebra methods are used for computing the eigensolutions and this part of the calculation becomes the bottleneck when the number of atoms reaches hundreds. On the other hand, at this regime of large system-sizes, the chemical interactions can be well approximated by localized basis sets, allowing a sparse matrix representation of the Hamiltonian and the overlap matrices.
We have developed and benchmarked a PETSc/SLEPc based sparse eigensolver that makes use of shift-and-invert parallel spectral transformations (SIPs) method. We demonstrate three main advantages of SIPs compared to dense solvers for the SCF method:
- SIPs exploits the sparsity of the matrices, hence reduces the memory footprint, and computational complexity significantly
- SIPs divides the eigenvalue problem into chunks that can be solved independently enabling scalability up to hundreds of thousands of cores with much less communications
- SIPs makes use of the eigenvalue distribution at a previous iteration to improve the job balance in the subsequent iteration allowing faster computations as SCF iterations converge
Through collaboration with this LDRD project, SIPs has become available through the SLEPc library and we have integrated this eigensolver into the SIESTA ab initio molecular dynamics package. We will present benchmark results for SIPs and SIESTA-SIPs on Cori, Mira and Theta supercomputers for different applications including carbon nanotubes, boron nitride doped graphene sheets and water clusters. Another aspect of our LDRD project was to perform benchmark coupled-cluster calculations based on the Feller-Peterson-Dixon method for the energetic properties of transition metal oxide (TMO) clusters. Such calculations are especially valuable to develop training sets for parameterized methods (density-functional tight-binding, semi-empirical molecular orbital) that yield sparse Hamiltonian matrices. We computed the heats of formation and the normalized clustering energies for the group 4 and group 6 TMOs and compared the accuracy of 55 exchange-correlation density functionals including pure and hybrid types.
Biography
Murat Keçeli obtained B.S. and M.S. degrees in physics from Bilkent University in Ankara, Turkey. After moving to the U.S., he joined Prof. So Hirata’s group, where he worked on the development of size-extensive vibrational self-consistent field theory, optimized vibrational coordinates, crystal-orbital modulo coupled-cluster and linear-scaling binary interaction methods for molecules and extended systems. He received his Ph.D. degree from University of Illinois at Urbana-Champaign in 2012 and joined Prof. Bill Green’s group at Massachusetts Institute of Technology as a postdoctoral associate. He worked on model reduction methods based on lumping analysis for large kinetic models and contributed to the development of a parallel quantum chemistry module for the reaction mechanism generator software. In 2014, Keçeli joined Argonne as a postdoctoral appointee supervised by Al Wagner.