30. Accelerating DMFT-MatDeLab GW Calculation with GPU
Authors: Zhihua Dong (Brookhaven National Laboratory)Kwangmin Yu (Brookhaven National Laboratory)Sangkook Choi (Brookhaven National Laboratory)Nicholas D'Imperio (Brookhaven National Laboratory)
Abstract: GW approximation and DMFT are theories to calculate electron structure in material science. The electron self-energy Σ and polarizability χ matrix elements construction in GW calculation iterations are the most time consuming parts, consuming about 40% of the total running time when calculating the Cerium α band on Titan at the ORNL Leadership Computing Facility. GPU acceleration was implemented on those two parts. To optimize the code, we overcome the difficulty of heavy non-sequential memory access caused by updating random elements of arrays at the inner most loop by sorting the index pattern and rearranging the calculation so that each GPU thread calculates one element, avoiding memory writing during updates. Test runs were conducted on Titan with 768 MPI processes. Results show a 26X speed up of the χ matrix calculation and 31.6% performance improvement of a complete α -Ce run.
Two-page extended abstract: pdf