CSM12 - Optimization of the Parallel Solver Execution Time with Evolutionary and Swarm Algorithms
Presenter
Event Type
Poster
TimeThursday, 13 June 201919:50 - 21:50
LocationHG EO Nord
DescriptionThe development of optimal designs requires access to a great computational power leading to the use of High Performance Computing. This work is focused on automatized settings of domain decomposition type solvers and algebraic multigrid solver available in ESPRESO, which is an open source computational tool for numerical simulations designed to utilize modern supercomputers. The optimal computational resource utilization requires good knowledge of the linear solver settings and might represent a problem for new users to achieve the best performance. We tried to improve this situation by an automatic setting of linear solver parameters with evolutionary and swarm optimization algorithms. We focused on a set of parameters which influences linear solver execution time, the most demanding part in simulation workflow. Especially in transient analysis, optimization algorithm is able to try, per each time step, different parameter combination representing an individual. With increasing number of processed time steps the execution time of the following time step decreases, as all individuals within the algorithm are reaching the best individual continuously.