Research On Fast(Parallel) Computing Of Disturbing Gravity

With the rapid increase of multi-source gravity observation data,it is necessary to give a pricise describe of gravity field and analyze the temporal-spatial distribution of gravity field from a global perspective.At present,the data processing technology of gravity field is not compatible with the development of gravity field observation technology,and the application of parallel computing technology in gravity field is not deep enough.Based on this,this paper taking the calculation of disturbing gravity as an example,the inherent parallelism of several methods to calculate gravity field parameters is analyzed,further experiments are conducted under different hardware conditions,especially in heterogeneous environment represented by CPU+GPU,to adapt to the requirements of larger-scale data processing.The main contents of this paper include1 In order to solve the problem of low efficiency of large linear equations when constructing the point mass model,the correlation between gravity anomaly and point mass is analyzed,and the solution scheme of residual point mass model based on window control is proposed.In this paper,the correlation of gravity anomaly is controlled within a certain range in the form of tierby-tier residuals.The coefficient matrix is converted from the original dense format to the sparse.The compressed storage can significantly reduce the amount of data storage.In the Intel Parallel Studio environment,a sparse linear equation of 112896 order can be solved efficiently by the MKL parallel computing library,which effectively enlarges the range of the point mass and improves the efficiency of the solution,result show that,on a multi-core processor,the computational efficiency can be further improved when using the MKL parallel computing library.2 Calculating gravity field parameters using gravity model is a time-consuming work especially for high-order gravity field model.Thus,a parallel computing scheme based on CUDA is designed in the GPU environment,which also applies to the calculation of other gravity field parameters.The calculation of disturbing gravity using gravity model can be divided into two parts: the calculation of Legendre function and its derivative and the calculation of spherical harmonic function.The former can only be completed serially,while the latter has abundant parallelism in computation.The calculation of the harmonic function is transformed into the point multiplication and dot product operation of the one-dimensional vector,and the corresponding GPU library function can greatly improve the efficiency of the calculation.With the increase of the computation order,the advantage of parallel computing will be more and more obvious.The parallel computing of disturbing gravity was demonstrated to outperform state of the art implementations using serial methods,just use one GPU card can helps to obtain a significant speedup.3 In order to achieve better performance,this paper optimizes the parallel computing scheme of model disturbing gravity based on the architecture features of GPU.It mainly includes the reasonable allocation and utilization of the threads,the optimization of data transmission between CPU and GPU,the optimization of memory access efficiency,the removal of conditional branches,and the reduction of computation time of serial code.In order to improve the efficiency of the ultra-high order Legendre function,the elements that are independent of the latitude and longitude are calculated in advance for repeated calls,which greatly save the time of serial parts.By removing the conditional branches,the Legendre function of its first derivative can be completed on the GPU,which is not only efficiency of the calculation,but also reduce the data copy to the device.The experimental results show that the above optimization can further improve the parallel efficiency,in the low-end consumer graphics device,when the order is greater than 1080,the overall computing performance increased by more than 5 times,in the Tesla K80 professional graphics card can improve at least 10 times.4 The parallel computing scheme of global model disturbing gravity based on 1D-FFT technique in GPU environment is designed.In this scheme,the calculation of spherical harmonics is completed on the GPU,and when the order is 2160,3 times(even higher)speedup can be achieved in the low-end consumer graphics card and more than 5 times in professional computing card.Based on the existing research results,the singularity-eliminated formula for calculating the horizontal component of disturbing gravity is presented.5 This paper realize the global integration of Stokes based on 1D-FFT to calculate the gravity field parameters for the first time.The biggest obstacle of global integration method is the timeconsuming calculating of Stokes kernel function,but the problem can be solved quickly after the parallelization.The experimental results show that when calculating the global disturbing gravity with resolution of 5 minutes,only less than 2 minutes was needed on Tesla K80,while the serial program will take about half an hour;when it comes to 2 arc-minutes,the serial algorithm takes about 12 hours,while the parallel algorithm only needs 20 minutes,more than 30 times faster that serial code.In order to further improve the performance,a parallel solution scheme for multimachine and multi-core heterogeneous environment is proposed.The program implements process-level parallelism between nodes,and achieves thread-level parallelism on multi-core in nodes which can fully exploit potential performance of different hardware in computer system.