(2018-01-23) Lena, Charles Manuel; Chelikowsky, James R.; Demkov, Alexander A; Ekerdt, John G; Hwang, Gyeong S; Gamba, Irene M

Show more

From the single hydrogen to proteins in the hundreds of thousands of kilodaltons, scientists can use the electronic structure of interacting atoms to predict their material properties. Knowing the material properties through solving the electronic structure problem, would allow for the controlled prediction and corresponding design of materials. The Kohn-Sham equations, based on density functional theory, transform a many-body problem impossible to solve for anything but the smallest molecules, into a practical problem which can be used to predict material properties. Although KSDFT scales as the cube of the number of electrons in the system, there are additional well documented approximations to further reduce the number of electrons, such as the pseudopotential method.
The incoming exascale era will lead to unavoidable challenges in solving the Kohn-Sham equations. These challenges include communication and hardware considerations. Old paradigms, epitomized by repeated series of globally forced synchronization points, will give way to new breeds of algorithms to maximizing scaling performance while maintaining portability.
This thesis focuses on the solution to Kohn-Sham DFT in real space at scale. Key to this effort is a parallel treatment of numerical elements involving the Rayleigh-Ritz method. At minimum, the Rayleigh-Ritz projection requires a number of distributed matrix vector operations equal to the number of electrons solved for in a system. Furthermore, the projection requires that number, squared and then halved, of dot products. The memory cost for such an algorithm also grows very large quickly, and explicit intelligent management is not an option. I demonstrate the computational requirements for the various steps in solving for the electronic structure problem for both large and small molecular systems. This thesis also discusses opportunities in real space Kohn-Sham DFT to further utilize floating point optimized hardware the with higher order stencils.