Hybrid Parallel Neighbor-related Computing For Vector Point And Raster Big Data Based On Heterogeneous CPU-GPU Systems

With the rapid development of data acquisition technologies,the data volume of spatial data has been increasing dramatically.Spatial big data poses a serious challenge to both computing hardware and algorithm design.As a typical type of computing,neighbor-related computing has been widely applied in many fields such as processing and analysis of raster and vector data,spatiotemporal process simulation and spatial statistical analysis.Faced with the requirement of fast processing of spatial big data,the performance of neighbor-related computing needs to be improved through parallelization.So for,the existing research mainly parallelizes neighbor-related computing based on multi-core CPU or many-core GPU.However,exploiting just one type of processing units can never make full use of the computing power of heterogeneous CPU-GPU systems.In order to further improve the efficiency of big data parallel processing,this paper studies the hybrid parallel neighbor-related computing based on heterogeneous CPU-GPU systems.To achieve highly efficient CPU-GPU collaboration,reasonable task mapping on CPU and GPU is firstly required.Considering the maturity of CPU parallelization and the complexity of GPU parallel design,this paper focus on GPU performance optimization for neighborrelated computing.On this basis,given the CPU-GPU performance gap and spatial heterogeneity,it is necessary to study the spatial decomposition strategy and task scheduling strategy so that load balancing can be achieved.Therefore,the main research contents and innovations of this paper include the following aspects:1)Faced with the challenges of data organization,parallel optimization and hardware adaptability,a GPU performance optimization model for neighbor-related computing is proposed.Specifically,the HG-GSPP model is proposed to organize vector points and raster data into a hierarchical grid,so as to achieve efficient use of GPU bandwidth.On this basis,a neighborhood sharing and task mapping model is proposed to optimize the parallel computing process.This model is based on neighborhood overlapping of adjacent target cells,and the caching of neighborhood data is thus realized by reutilizing register resources,which ultimately improves the efficiency of parallel computing.In terms of hardware adaptability,we take hardware capability,algorithm requirements and optimization rules into account,and propose a resource configuration optimization model based on dynamic programming to achieve hardware adaptiveness and efficient use of computing resources.2)Considering the CPU-GPU performance gap and spatial heterogeneity,it is difficult to achieve load balancing by simply using common load balancing strategies for hybrid parallel computing.Therefore,this paper firstly proposes the LogarithmicRPE model to fit the change trend of GPU performance.Applying this model,the performance of CPU and GPU can be estimated,and the optimal spatial decomposition granularity can be predicted.Based on the Logarithmic-RPE model,a hierarchical domain decomposition method and a line-scanning domain decomposition method are proposed for hybrid parallel computing.On this basis,the HS-TP scheduling strategy quantifies spatial heterogeneity by introducing task priorities so as to achieve load balancing and performance maximization.3)The existing research of parallel neighbor-related computing exploits only one type of processing units,thus resulting in resource underutilization in the CPU-GPU heterogeneous system.In order to support the partition and management of spatial big data in heterogeneous CPU-GPU systems and achieve highly efficient parallelization of neighbor-related computing,this paper proposes a hybrid parallel neighbor-related computing framework(HP-NRCF).In HP-NRCF,the data management module supports the expression,organization,decomposition and management of vector points and raster data;the computing engine module is responsible for computational resource management and efficient parallelization of neighbor-related calculations,and supports multiple parallel execution modes.The scheduling and algorithm library module provides efficient scheduling strategies and optimized algorithms for parallel neighbor-related computing.