Most computational science and engineering problems in domains such as computational fluid dynamics, thermodynamics, and electronics require solving large, sparse linear systems derived from discretizing partial differential equations. Due to the highly computationally expensive nature, many numerical algorithms and libraries implementing these algorithms have been developed through decades of research. PETSc (Portable Extensible Toolkit for Scientific Computing) is a widely used library due to the large pool of numer- ical solvers and preconditioners and the support for the rapidly advancing computing platforms. However, selecting the most suitable solver and preconditioner for solving the linear system is not straightforward. It would need extensive knowledge and experience in numerical mathematics, domain expertise, and reading through many literature. Attempts have been made to ease the process using machine learning techniques, which this study tries to follow to find the best solver-preconditioner pair for the pressure Poisson solver of the MGLET computational fluid dynamics code and extend the finding to heterogenous computing architectures involving GPUs.     «

Most computational science and engineering problems in domains such as computational fluid dynamics, thermodynamics, and electronics require solving large, sparse linear systems derived from discretizing partial differential equations. Due to the highly computationally expensive nature, many numerical algorithms and libraries implementing these algorithms have been developed through decades of research. PETSc (Portable Extensible Toolkit for Scientific Computing) is a widely used library due to...     »