High-Level Framework for Solving Systems of the PDEs on Distributed Systems

This work aims to develop a framework for the high-level and user-transparent distribution of calculations for solving systems of partial differential equations. Framework provides a simple high-level application programming interface and automatically partitions problems into sub-problems that are then executed on MPI-based clusters. Disposing of CUDA accelerators is supported. The management of processes and accelerators is performed by the framework using the MPI services. The performance of the developed system was analyzed, the influence of the data distribution scheme on performance was investigated, and the performance of the CPU and GPU modes of the framework was compared. Computational fluid dynamics problems on the 2D unstructured mesh were used as a benchmark task for the framework. Linear acceleration on a number of available CPUs is observed, with several super-linear cases.