Study On Execution Time Prediction For Parallel Geocomputation In Multi-Core Cluster Environment

The development of geocomputation impels the combination of the parallel computing and geography. Solving some high-complexity problems in geography by parallel computing becomes a developing trend. Last-model hardware architectures, as the multi-core processor and the cluster, break through the bottleneck of the massive data processing and complicated geographic space analyzing and modeling by the traditional GIS. It is a new researching approach for solving the problems of geocomputation. The study on the execution time prediction for the parallel geocomputation, as one of the most important technologies of the parallel computing, will bring great real meaning and scientific value for the task scheduling and the load balancing. This dissertation focuses on the execution time prediction for the parallel geocomputation and study on the following aspects.Firstly, this dissertation proposes and models the execution time of OpenMP parallel algorithms and MPI parallel algorithms by the static analyzing method, according to the characteristics of different parallel programming models. Analyzing the factors which have effect on the execution time and its changing trend, which are the theoretic support of the dynamic prediction.Secondly, based on the raster data, this dissertation proposes the factors which have effect on the scale of the raster data, which provides the evidence for the parameter of the prediction model. According to the dynamic predicting, using the interpolation and polyfit methods to fit a three-dimensional prediction model with the parameters of the data scale, the processor number and the parallel running time, which are the basement of the task scheduling. This method can acquire higher accuracy.Finally, this dissertation integrates Torque and Maui into a task management configured on the current cluster. Scheduling a group of tasks with the backfill algorithm by using the prediction execution time of each task as one of the input parameters. Compare with the FCFS algorithm and the backfill algorithm with no prediction execution time parameter, the load of the system resources can reach the better performance by the backfill algorithm with the prediction execution time parameter. The experiments validate that all execution time prediction methods and solutions in this dissertation is accuracy, feasible, dependable and effective and show that the execution time prediction technique can effectively instruct the task scheduling and load balancing.