【作者】 张帆；

【导师】 谭捍东；

【作者基本信息】 中国地质大学（北京） ， 地球探测与信息技术， 2011， 博士

【摘要】 电磁法勘探是勘探地球物理方法中重要的分支,电磁法勘探分为两大类,即传导类电磁法和感应类电磁法,传导类电磁法中是以直流电法为主;感应类电磁法中,主要是电磁测深法,其中具有代表性的是大地电磁法。随着电磁法勘探技术的不断发展,特别是三维正反演技术的不断更新,大规模计算问题也越来越多,迫切需要将并行计算最新技术应用到电磁法勘探解释工作中来。直流电法和大地电磁法的三维正演都存在需要求解大型稀疏对称系数矩阵线性方程组的问题,计算消耗的时间很长,降低了正演的运算效率,同时降低了三维反演的效率。因为直流电法和大地电磁法的三维正演计算问题都具备一定的并行性,本论文采用并行计算最新技术基于GPU并行计算和基于MPI的并行计算相结合的计算模式对这两种电磁法已有的三维正演串行算法进行并行化处理。即基于MPI的并行方法粗粒度的分配计算任务到各个计算节点,然后基于GPU的并行计算将计算任务繁重的大型稀疏对称系数矩阵线性方程组的求解进行细粒度的并行化处理,使用GPU的众多流处理器执行运算。对于直流电法,将供电电源多次移动,即多个供电电源的位置上测得的测量电极电位差的多个数据集粗粒度的分配到各个计算节点,每个计算节点在进行三维正演计算时将基于共轭梯度算法的求解大型稀疏对称系数矩阵线性方程组的过程细粒度并行化,分解到GPU的各个流处理器中执行运算。对于大地电磁法,因为所处理的数据是按照不相同的多个频率来处理的,所以将大地电磁三维正演算法所需要计算的频率基于MPI并行粗粒度的分配到各个计算节点中,在每个节点在进行三维正演计算时将基于共轭梯度算法的求解大型稀疏对称系数矩阵线性方程组的过程细粒度并行化,分解到GPU的各个流处理器中执行运算。最后,选取不同模型,通过实验分析验证了直流电法和大地电磁的三维正演并行算法的正确性和性能。通过该并行计算模式能够实现更大限度的并行化,更好的提高这两种正演算法的运算效率。更多还原

【Abstract】 Electromagetic exploration is an important branch of exploration geophysics methods and has two categories. One is conductive electromagnetic method; another one is inductive electromagnetic method. The main method of conductive electromagnetic method is direct currency method. The main method of inductive electromagnetic method is electromagetic sounding method, which is representative of magnetotelluric method. With the development of electromagnetic exploration technology, especially the update of three-dimensional forward modeling and inversion technology, there are more and more large-scale computation problems. So it urgently needs to apply the latest technology of parallel computing to the interpretation of electromagnetic exploration.Both of three-dimensional forward modeling of DC method and magnetotelluric method need to solve linear equations with large sparse symmetric coefficient matrix. It takes long computation time and reduces the computation efficiency for forward modeling, while reducing the efficiency of three-dimensional inversion. Because there is parallelism in both of three-dimensional forward modeling computation of DC method and magnetotelluric method, this topic uses the latest technology of parallel computing based on GPU combined with parallel computing based on MPI to parallelize traditional three-dimensional forward modeling serial algorithms of DC method and magnetotelluric method. Parallel computing based on MPI coarse-grainedly distributes computing tasks to each computing node, and then each computing node use parallel computing based on GPU to fine-grainedly solve linear equations with large sparse symmetric coefficient matrix. Let many stream processors of GPU execute heavy computing tasks.For DC method, parallel computing based on MPI coarse-grainedly distributes several data sets collected by measuring electrode based on moving point source which is equivalent to several different positions of point source to each computing node, and then each computing node fine-grainedly parallelizes the process of solving linear equations with large sparse symmetric coefficient matrix using conjugate gradient method, distributing to many stream processors of GPU to execute computing tasks.For magnetotelluric method, because all the processed data are based on a number of different frequencies, so parallel computing based on MPI coarse-grainedly distributes all these frequencies which are for processing data to each computing node, and then each computing node fine-grainedly parallelizes the method of solving linear equations with large sparse symmetric coefficient matrix using conjugate gradient method, distributing to many stream processors of GPU to execute computing tasks.Finally, we verify the validity and performance of both of the three-dimensional parallel algorithms of DC method and magnetotelluric method by experiments choosing different models. Using this new parallel computing mode can achieve a greater level of parallelism and considerably improve both of three dimensional forward modeling algorithms of DC method and magnetotelluric method.更多还原