Agents Based Hierarchical Parallelization of Complex Algorithms on the Example of hp Finite Element Method

The paper presents how application of agents can improve scalability of domain decomposition (DD) based parallel codes, where the optimal load balance for some components of the code cannot be achieved only by partitioning computational domain. The limitation of the DD paradigm, where some highly overloaded pieces of domain cannot be partitioned into smaller sub-domains can be effectively overcome by parallelization of computational algorithm over these pieces. The agents are used to localize such highly loaded unbreakable parts of domain. Multiple agents are then assign to each highly loaded part to execute computational algorithm in parallel. The resulting hierarchical parallelization scheme results in the significant improvement of the scalability. The proposed agent based hierarchical parallelization scheme has been successfully tested on a very complex hp Finite Element Method (FEM) parallel code, applied for simulating Step-and-Flash-Imprint Lithography (SFIL), resistance heating of Al-Si billet in steel die for tixoforming process as well as for the Fichera model problem.