Topic/Type: 2.5 Adaptative & multi-scale methods, Poster

A multi-scale electromagnetic particle code with adaptive mesh refinement and its parallelization

M. Nunami1, 5, H. Usui2, 5, Y. Kajimura3, 5, T. Moritaka2, 5, I. Shinohara4, 5

1 National Institute for Fusion Science
2 Kobe University
3 Kyoto University
4 Japan Aerospace Exploration Agency
5 Japan Science and Technology Agency, CREST

We have developed a new electromagnetic particle code with adaptive mesh refinement (AMR) technique which is one of the most powerful tools to analyze the multi-scale plasma physics. AMR can realize high-resolution calculation saving computer resources by generating and removing hierarchical cells dynamically according to certain refinement criteria, monitoring cells where high-resolution are needed.
In the developed code, we have applied AMR to electromagnetic particle-in-cell (PIC) method by using fully threaded tree data structure [1] and have parallelized the code with load balancing.
In the parallelization, we adopt domain decomposition method, and for good locality preserving and dynamical load balancing, we use Morton ordered (Z-ordered) curve [2] which is a kind of space-filling curve. In PIC, particle calculation occupies most of total calculation time. Furthermore, in AMR, not only width of cells but also width of time steps are refined. For load balancing in PIC, it is the most essential to consider the number of particle calculation loops for each cell among all hierarchical levels as a work weight for each processor. Therefore, we concentrate work weights based on the number of particles and hierarchical level of each cell. Then we distribute the domain according to Motron curve and the work weight, so that each processor has approximately the same amount of work.

[1] Khokhlov, J. Comput. Phys. 143, 519 (1998).

[2] M. J. Warren & J. K. Salmon, in Proc. Supercomputing (Washington DC: IEEE Comput. Soc.), 12 (1993).