Hybrid GPU/CPU Approach to Multiphysics Simulation

Multiphysics simulation is a crucial stage in many current fields of science and industry like solid mechanics, heat transfer, automobile systems, or nuclear systems. Modeling and simulation provides knowledge that would be difficult or even impossible to obtain when trying to experiment with physical objects. Multiphysics simulation is a time and computational demanding process. Therefore great effort has been put forward to apply new technologies which eventually will enhance performance. In the last decade a number of approaches were proposed. One approach takes advantage of classic supercomputers with thousands of Central Processing Units (CPU). Another approach uses modern Graphical Processing Units (GPU) to perform general purpose computing (GPGPU) by adjusting the application to a new environment. In this paper, hybrid approaches that use both CPU and GPU are proposed. The approach assumes executing highly serial parts of code by a classic multiprocessor machine that uses underlying GPU to execute heavy computational and parallel parts of code. The idea is applied to a Finite Element (FE) library called libMesh (a part of the MOOSE Framework). Proposed modification of the FE library does not affect existing applications that use the MOOSE Framework or libMesh alone. Therefore, it is enough to recompile the multiphysics simulation framework of the library. To evaluate performance of implemented modification, software was used to perform a simulation on an empty cylinder sealed from both sides. Obtained results show that the presented approach has potential and may be beneficial to develop this idea by extending the scope of the code executed by GPU.