Chronos: an AMG solver for numerical simulations on HPC platforms

Nowadays, numerical simulation has become a key element in solving problems of engineering and industrial interest. In fact, the need to improve the accuracy of the solution drives the development of more complex numerical models with a high degree of spatial discretization. On large complex problems, the most demanding phase of the simulation is the solution of the system of linear equations derived from the discretization of partial differential equations governing physical process. The solution of these systems requires the use of infrastructures designed for high performance computing (HPC) which, through the development of cloud platforms have become easily accessible for all engineers. This thesis work presents a general-purpose Algebraic MultiGrid (AMG) linear solver designed for distributed memory HPC systems equipped with GPU accelerators. The novelty of this research is the development and implementation of algorithms based on known numerical approaches, capable of exploiting the hardware resources of today’s HPC systems. In particular, the work focuses on the redesign of algorithms developed for multi-core CPUs and their porting to GPU boards. This research project is co-funded by M3E [M3E, 2023], and the developments are integrated into the proprietary software Chronos. The effectiveness and performance of the proposed solver have been validated by solving a large set of problems arising from real-world applications on the Marconi100 supercomputer.