Research Of Parallel Statistical Cost Flow Phase Unwrapping Algorithm Based On Network Flow Model

With the increasingly wide range of computer applications and the growing scaleof the processing problem, the computer hardware has been developing rapidly. Inrecent years, it has entered into a stage of development, which has multi-corearchitecture, personal high-performance computer, and petaflops parallel machine. Inorder to adapt to the rapid development of computer hardware and meet variousapplication fields’ growing surge of demand for computing power, parallel computingtechnology has been developed rapidly and is widely used in astrophysics, fluiddynamics, reservoir modeling, ocean circulation, and long-term weather forecasting,seismic data processing, biological information processing, computer-aided design,database management, image processing, as well as commercial search engines andother fields.Parallel computing is divided into distributed parallel computing and sharingstorage parallel computing in line with different processor architecture and storage type.Generally speaking, distributed parallel computing adopts message passingmodel.MPI(Message Passing Interface) is the current message passing programmingstandard, which has the advantage of good portability, high efficiency and so on.However, OpenMP is the standard of sharing storage model, which is more effective inthe description of communication between processors within single SMP node. Atpresent, the MPI+OpenMP multi-level parallel programming model is the combinationof these two advantages, has become a mainstream programming technology. It cangive full play to complementary advantages in today’s multi-core architecture.The application foundation in this thesis is statistical cost flow phase unwrappingalgorithm based on network which is the key technique of InSAR (InterferometricSynthetic Aperture Radar) technology. Aim at the problems when processing massiveSAR data, such as the shortage of computing resource and the low relative computingefficiency, the parallelization of this algorithm is proposed in this research, in order toenhance the computing efficiency of phase unwrapping, so that it will be helpful forother link in InSAR data processing.Firstly, this thesis studies the domestic and international parallel computing technology as well as some classic parallel programming model in recent decades.Then, the MPI and OpenMP are chosen as the main study objects, analysis theirprogramming models and the MPI+OpenMP multi-level parallel programming model.On that basis, analysis and research the serial statistical cost flow phase unwrappingalgorithm, excavate its inherent parallelism, design and realize its parallel algorithmbased on MPI. Further, analysis the computing code of the fine granularity in thisserial algorithm, and implement its parallel algorithm on account of MPI+OpenMP. Inthe course of the study, the different-calculating-scale performance comparativeexperiment of the parallel algorithm under the environment of MPI andMPI+OpenMPis respectively carried out, so as to test and verify the multi-level parallelprogramming has better acceleration performance for parallel computing.During the process to realize the parallel statistical cost flow phase unwrappingalgorithm based on MPI, the parallel I/O technology in MPI-2is studied as well.According to I/O operational characteristics in the course of phase unwrapping, thedisplay-offset parallel I/O pattern is applied in statistical cost flow phase unwrappingalgorithm. A better performance of this algorithm has been proved by means ofexperiment when adding parallel I/O technology.The parallel statistical cost flow phase unwrapping algorithm based on pure MPIand multi-level parallel MPI+OpenMP both enhanced phase unwrapping efficiencyand reduced the amount of memory used in the course of unwrapping, throughexperiments and analysis, and the performance of MPI+OpenMP multi-level parallelalgorithm is better than pure MPI parallel algorithm under the same condition. Inaddition, the application of parallel I/O technology in this algorithm has reduced thetime of reading and writing files, which has further increased the efficiency of parallelalgorithm.