GPU Acceleration of a Global Atmospheric Model using Python based Multi-platform

A global atmospheric model play an important role in short-term weather forecasting and long-term climate prediction. The model requires enormous computing resources because the all atmospheric states must be calculated every time step (usually a tens of seconds to several minutes). However, since the most atmospheric models run only on CPU machines, they are not able to use the modern microprocessors with high performance and low power such as NVIDIA GPU and Intel MIC.

It often costs a lot to convert codes from one machine to the other machine. Although it can be accelerated on GPU and MIC using OpenMP and OpenACC directives, it is not easy to achieve peak performance. I developed a new Python module named PyMIP (Python based Machine Independent Platform) to integrate C, Fortran, CUDA and OpenCL codes with a simple user interface. The main code includes configuration, flow control, IO and MPI parallel is written by Python. Only hotspots include huge number crunching code are written by compile language as C, Fortran, CUDA and OpenCL. The hotspot codes are compiled and imported using PyMIP in runtime. PyMIP enables that a user can switch machines with simple flag.

I am developing a new global atmospheric model based on PyMIP to make it easy to utilize various modern microprocessors. In this presentation, I will introduce PyMIP and show the computational performance result in NVIDIA GPU of the dynamical core of the model developed based on PyMIP.