Fresnel层析成像并行算法研究

与射线层析成像相比,Fresnel层析成像考虑波频率的影响,具有较高的分辨率,但所需的存储空间和计算量更大,因此提出了Fresnel层析成像的并行算法.把大型层析反演方程组的求解,转化成对其中的各个方程进行相互独立的计算,避免了大型系数矩阵的存储问题;把一个Fresnel带的正演和反演计算放在一个进程,不同Fresnel带的计算相互独立进行,不需要信息传递,达到了极高的并行度;从进程之间没有通信,仅当从进程计算结束后,在主进程与各从进程之间有少量的数据传递,使通信开销达到了极小的程度.应用MPI在LinuxPC集群环境下实现了该算法,实际测试表明,该算法具有较高的并行度和加速比.

A Parallel Algorithm for Fresnel Tomography

In contrast with ray-based traveltime tomography,Fresnel tomography accounts for the band-limited nature of seismic waves and gives the higher resolution tomograms.Because Fresnel tomography demands much computer memory and much running time,a parallel algorithm for it is proposed.The tomographic inversion is transformed to resolving respectively a series of single equation in light of backprojection principle,each equation corresponding to a Fresnel zone.The forward and inverse computation concerning a Fresnel zone is allocated to one process and is independent of other processes.Then the storage and calculation of the large-scale matrix in the tomography are avoided.No message delivers between the slave processes,and only a little of data delivers between a master process and the slave ones.By using the portable message passing interface standard(MPI) for the communication,the computing code of the algorithm is implemented on Linux system,which allows to distribute the work on several PCs connected via standard Ethernet in an in-house network,and greatly expands the applicability of Fresnel tomography.The tests on the synthetic and observed seismic travel time data show that this parallel algorithm has a good performance on Linux PCs.