Dynamic load balancing with the parallel partitioning tool GridSpiderPar

Dynamically adaptive meshes are used often in high-performance computing. A mesh is locally refined or de-refined in spots of interest or where high gradients of an objective function arise. Its repartitioning is needed periodically to balance the load on the processors. Dynamic load balancing algorithms were developed on the base of the parallel geometric algorithm of mesh partitioning and the parallel incremental algorithm of graph partitioning from the partitioning tool GridSpiderPar. Initial partition of a mesh with local refinement (6.7$\cdot$10$^6$ hexahedrons) was compared with the results of repartitioning using the devised algorithms. Comparison results show advantages of the parallel geometric algorithm on this mesh and features of the parallel incremental algorithm application.