Acceleration Of Particle Transport Simulation Based On High Performance Coprocessors

The particle transport equation describes the conservation relative for mass,charge,momentum and energy between the particles in some medium mathematically.A solution of the particle transport equation plays a fundamental role in many different physical and engineering fields,such as celestial objects detection,radiation therapy in medicine,nuclear reactors and nuclear weapons design.Now,many applications have increased the need for high fidelity particle transport simulations,and then the more refined physical models have resulted in the sharp increase of computational scales.In addition,some specific applications demand highly real-time solutions.In recent years,the application of co-processors with high performance-power efficiencies is an important trend in high performance parallel computing.Thus the use of different co-processors is becoming a promising solution for accelerating particle transport simulations.However,there are some challenges: parallel algorithm design and optimization,the diversity of programming models,and the determination of appropriate co-processors.In order to meet these challenges,this thesis emphasizes on the parallel algorithms and implementations of both deterministic and Monte Carlo particle transport numerical methods on MIC and FPGA co-processers.In summary,this thesis makes the following contributions:1.This thesis proposes a MIC-based three-dimensional structured grid multi-level parallel sweeping algorithm for the parallel solution of particle transport finite difference discrete ordinates equations in structured grids.The algorithm exploits the multi-level parallelism in wavefront sweeping,maps I-line grid columns in a wavefront to parallel threads in MIC,and vectorizes the iterative solution of finite difference discrete ordinates equations in a I-line grid column by isolating the computation of some physical variables.The results show that MIC gets 2.03 times speedup compared with multi-core CPU when flux fixup is off.With flux fixup,MIC can improve performance with a factor of 1.50 times in comparison with multi-core CPU.2.This thesis proposes two MIC-based two-dimensional unstructured grid multi-level parallel sweeping algorithms for the parallel solution of particle transport discontinuous finite element discrete ordinates equations in unstructured grids.A search algorithm is introduced to determine the cells and the order of wavefronts before the parallel solution.One algorithm exploits the thread-level parallelism in wavefront sweeps of all energy groups with hardware threads and the data-level parallelism in the single energy-group solution of one cell with vector units.The other cooperatively exploits the parallelism in wavefront sweeps of all energy groups with hardware threads and vector units,and adopts the memory optimization.The results show that two algorithms on MIC can obtain the speedup of 39.92 times and71.54 times compared with the serial on CPU,respectively.3.This thesis proposes a MIC-based fast Monte Carlo particle transport multi-level parallel simulation algorithm for the parallel solution of coupled electron-photon fast Monte Carlo transport dose simulation(DPM).The algorithm introduces a new multi-level parallel data structure to satisfy the memory demand of vector-level and thread-level parallelization,involves the thread-level parallelization and the data locality optimization,and implements the vectorization of electron transport simulation by developing a multi-level parallel random number generator.The results show that the parallel solution on MIC can exhibit very accurate results and achieve the speedup of 16.22 to 18.82 times in comparison with the original DPM serial code on CPU.4.This thesis proposes an FPGA-based fast Monte Carlo simulation architecture for the fast solution of photon fast Monte Carlo transport dose simulation.This structure is based on single precision floating point,and has good performance on speed and power consumption by implementing pipeline parallelization,bit-level parallelization and other special designs.The results show that the simulation results of the structure are in good agreement with the results from the software on CPU.Compared with the serial implementations on 2.30 Ghz CPU and 3.40 GHz CPU,the structure can get the speedup of 22.15 times and 33.18 times,respectively.In summary,this thesis deeply studies the parallel solutions of particle transport numerical simulation on MIC and FPGA co-processers so as to effectively exploit the huge computational power of MIC and FPGA co-processors.The study lays the foundation for the application of some numerical simulations in practice,and the construction and application of co-processers based massively parallel computing dedicated systems,which service massively parallel simulations of particle transport.