98. First Experiences with ab initio Molecular Dynamics on OpenPOWER: The Case of CPMD
Authors: Valery Weber (IBM)A. Cristiano I. Malossi (IBM)Ivano Tavernelli (IBM)Teodoro Laino (IBM)Costas Bekas (IBM)Manish Modani (IBM)Nina Wilner (IBM)Tom Heller (IBM)Alessandro Curioni (IBM)
Abstract: The algorithmic adaptation and code re-engineering to port CPMD code to next-generation heterogeneous Open-POWER architectures (CPU+GPU) is presented here. The construction of the electronic density, the application of the potential to the wavefunctions, and the orthogonalization procedure are offloaded to the GPU. The different GPU kernels consist mainly of fast Fourier transforms (FFT) and basic linear algebra operations (BLAS). The performance of the new implementation obtained on Firestone (POWER8/Tesla) is discussed. The communication between the host and the GPU contributes a large fraction of the total run time. We expect a strong attenuation of the communication bottleneck when the NVLink high-speed inter-connect, between CPU to GPU will be available. These results will be in the final version of the poster for SC16.
Two-page extended abstract: pdf