B44.00005. Recent Advances in the Development of the Octopus Code for GPUs
Presented by: Sebastian Ohlmann
Since the end of Moore's law, using state-of-the-art supercomputers has become more and more difficult because simulation codes need to expose more and more parallelism to benefit from larger machines. Moreover, supercomputers have become heterogeneous as the increase in CPU performance has slowed down: all of the first three exascale machines to be built in the US will heavily rely on GPUs to achieve their compute performance. Thus, simulation codes in all areas, also real-space DFT codes, will need to utilize GPUs efficiently in order to use these next-generation machines. In this contribution, we report on recent advances in the development of the real-space DFT code Octopus for GPUs. Those developments include porting more algorithms to GPUs, both for ground-state and time-dependent calculations, improving the data handling when copying data to and from the GPU and improving the overlap of communication with computation. We will show how these advances improve the speed of the code and its scalability and apply the code to some larger problem sizes on GPUs.
- Sebastian Ohlmann