Heterogeneous Parallel Method For SPH-centric Fluid Dynamics

In the recent decade fluid dynamics employed in smooth particle hydrodynamics(SPH)methods are gaining popularity among researchers in many relevant fields.At present,larger scale of fluid simulations with more details,more accuracy,and faster speed are urgently needed.Yet,such requirements would demand more particles and higher particle density that unavoid-ably would increase calculation significantly.However,the existing SPH implementations based on general-purpose computing on graphics processing units(GPGPU)still can hardly meet the demand for the huge amount of calculation.In this paper,a CPU-GPU heterogeneous architecture which is designed to minimize the overall cost of the previous shared memory based GPGPU implementation for the SPH methods will be introduced.This architecture uses CPU to share certain workloads which are otherwise relatively inefficient for GPU.For more efficient cooperation between CPU and GPU,the cus-tomized task assignment policy,load balancing strategy and data exchange method are included in this architecture.An expected improvement can be achieved with the help of this architecture.The previous GPGPU implementations for the SPH method can hardly fully dig out the performance of GPGPU.In this paper,a novel GPGPU system-level framework for the SPH method will be introduced.This framework includes a newly designed task assignment algo-rithm and particle interaction method.They can unify several characteristics of GPGPU with the advantages of the underlying SPH algorithms in a seamless way.This framework can exhibit up to 60%performance improvement when compared with other state-of-the-art methods.The novel GPGPU system-level framework for the SPH method is enabled by certain pow-erful technologies and methodologies available during the analysis of GPGPU program perfor-mance.Such analysis method will be introduced in this paper.A theoretical and comprehensive analysis and comparison between the existing works will also be done with the help of such analysis method for explicitly identifying their differences.