基于大地电磁二维反演的MPI和CUDA并行算法研究

A Study of MPI and CUDA Parallel Computation Based on MT Inversion Algorithm

大地电磁测深法属于天然场源的电磁勘探方法,是以岩石的电性差异为基础和前提的勘探方法。所采用的大地电磁二维反演方法为共轭梯度法,该方法避免了求解雅可比矩阵,效率较高,但是在将模型剖分较细时,多频率进行计算时效率有待提高。基于大地电磁频率依次独立处理数据的特点,采用了MPI的并行算法来提高效率,为了使得计算效率更高,在MPI的基础上增加了CUDA并行运算方法,用多个进程同时来计算各频率数据,在求解方程的过程中采用CUDA进行加速,得到计算结果。通过对正演和反演图的比较,验证了程序的正确性。对并行算法的效率进行了统计,进程数为2~4时,加速比能达到2.15~3.09,比单一的MPI并行算法的加速比要高,验证了程序的有效性。

Magnetotelluric sounding method based on the difference of the rock＇s resistivity is an Exploration method which do research in earth’s resistivity and phase using the native electromagnetic field. The inversion algo-rithm of 2D magnetotelluric named conjugate gradient method was intraduced, the method which avoid resolving the Jacobi matrix very effective, but the method is not effective enough when the model is divided into big grid. Based on the characteristics of the magnetotelluric processing data according to frequency independently, MPI technology was applied in the algorithm. In order to increase processing efficiency, CUDA technology was added to the algo-rithm, and the time of solving equation is less than the previous time. After comparison of the forward and inversion result, the algorithm is proved aright. The papaer has recorded the efficiency of the algorithm. When the number of the processes is 2 to 4, the speedup ratio is 2. 15 to 3. 09. the efficiency of the mixed parallel algorithm is higher the MPI algorithm.