Acceleration Of Non-equilibrium Ionization Simulation

Astrophysical simulations are computation-intensive and time-consuming,so modern astrophysical codes aim at a good balance between accuracy and performance in massive parallel environments.Meanwhile,non-equilibrium ionization(NEI)is an important phenomenon related to various astrophysical processes,and its full modeling and simulation is still one of the biggest challenges faced by modern astrophysics.However the traditional time-splitting scheme tightly coupled the NEI solver with the underlying Eulerian mesh and introduced high overhead on computing,memory and communication.In this thesis,in order to accelerate NEI calculation on multi-core heterogeneous systems and large-scale parallel environments,several optimization schemes have been designed and implemented at three levels of simulation process,computing framework,and numerical algorithm.Firstly,the performance bottlenecks of the traditional NEI scheme have been explained and validated.By introducing tracer particles the NEI solver is decoupled from underlying Eulerian mesh,and based on MapReduce model,a full parallel pipeline is proposed to support both post-processing and nonintrusive simulation-time data analysis.In order to quickly process large amounts of small particle snapshots continuously produced during the course of simulation,several I/O optimization schemes are also designed,including serial I/O mode,direct I/O mode,and real-time stream mode.Evaluations on up to 192 cores show that our approach can improve the end-to-end performance of a real world simulation by 3-fold above.Secondly,in order to break through the performance limitation of the traditional homogeneous parallel processing systems based on CPUs,which is not efficient to tackle numerous compute-intensive tasks,a GPU-optimized NEI ODE solver is designed based on the CUDA programming model.Moreover a load balance strategy on hybrid multiple CPUs and GPUs architecture via share memory is also proposed to maximize performance and device utilization.Comparing with the 24 CPU cores(2.5GHz)parallel version,this implementation on 3 Tesla C2075 GPUs achieves a speed-up of up to 15.Thirdly,in order to improve the traditional non-interactive mode of running simulations,a framework for adding interactive control functionalities during the course of time-consuming astrophysical simulations is presented,which is based on a hierarchical architecture where computational steering in the high-resolution run is performed under the guide of knowledge obtained in the gradually refined ensemble analyses.Several visualization schemes for facilitating ensemble management,error analysis,parameter grouping and tuning are also integrated owing to the pluggable modular design.The visual steering approach effectively accelerates the process of simulation modeling and ionization state analysis,and facilitates decision making on parameter tuning and numerical error control etc.Additionally,the thesis is a result of three-year collaboration with astronomers.All the proposed schemes are prototyped based on the FLASH multiphysics simulation framework and evaluated by real-world simulations,i.e.,supernova remnant,nucleosynthesis,spectral calculation and stellar wind.The accuracy of all the proposed schemes is also validated and confirmed by domain experts.Moreover the adaptability of these methods to other reactive flow simulations is also explored and validated.The optimization approaches for NEI calculation issued in this thesis have high reference values for resolving other large-scale astrophysical simulation in future.