5:15 PM - 6:45 PM
[PEM17-P02] Scalable Domain-Decomposing Dynamic Load Balancing for Planetary Upper Atmosphere DSMC Model
Keywords:Planetary Upper Atmosphere, Planetary Exosphere, Direct Simulation Monte Carlo, Dynamic Load Balancing, Parallel Simulations
The escaping processes of planetary upper atmosphere is a key scientific issue in understanding the atmospheric evolution of planets such as Mars. The Direct Simulation Monte Carlo (DSMC) model has been developed for the dilute upper atmosphere of Mars [Terada et al., 2016], where the intermolecular collisions are not so frequent enough to be treated as a fluid, but the collision process still plays an important role. In the DSMC method, the intermolecular collisions are processed for particles residing in the same computational cell. The developed model has successfully reproduced thermal and non-thermal escapes in a self-consistent manner, taking in place in the upper atmosphere of terrestrial planets. To estimate the actual atmospheric outflow rate and its dependence on the solar EUV radiation, however, it is necessary to further improve the computational efficiency on a large distributed-memory computer system.
The fundamental parallelization strategy for particle simulations can be categorized into the particle splitting and domain splitting methods. In the DSMC method, which processes the inter-particle collisions, it is difficult to arbitrarily split the particle population and assign the subset of particles to different MPI processes. Thus, it is natural to adopt the domain decomposition method. On the other hand, the particle load imbalance among subdomains caused by the inhomogeneity of particle density tends to degrade parallel performance of the domain-decomposing simulations. The objective of this study is to realize highly scalable parallel DSMC simulations by applying the OhHelp algorithm [Nakashima et al., 2009]. OhHelp is a dynamic load balancing technique originally designed for plasma particle simulations based on the particle-in-cell method. Load balancing in terms of the particle populations are accomplished by making each MPI process help another process having a densely populated subdomain. The presentation will review the basic algorithm of the OhHelp method and discuss necessary considerations for its implementation in DSMC computations. Preliminary results of parallel performance evaluation will also be presented.
The fundamental parallelization strategy for particle simulations can be categorized into the particle splitting and domain splitting methods. In the DSMC method, which processes the inter-particle collisions, it is difficult to arbitrarily split the particle population and assign the subset of particles to different MPI processes. Thus, it is natural to adopt the domain decomposition method. On the other hand, the particle load imbalance among subdomains caused by the inhomogeneity of particle density tends to degrade parallel performance of the domain-decomposing simulations. The objective of this study is to realize highly scalable parallel DSMC simulations by applying the OhHelp algorithm [Nakashima et al., 2009]. OhHelp is a dynamic load balancing technique originally designed for plasma particle simulations based on the particle-in-cell method. Load balancing in terms of the particle populations are accomplished by making each MPI process help another process having a densely populated subdomain. The presentation will review the basic algorithm of the OhHelp method and discuss necessary considerations for its implementation in DSMC computations. Preliminary results of parallel performance evaluation will also be presented.