The Parallelization Research Of Map Algebra Of Spacial Analysis In Heterogeneous Computing Environment

All along, the goal of spatial analysis researchers is how to quickly extract richer and more useful information from the spatial data to provide reference for decision by organizing and using the spatial data efficiently. With the continuous improvement of the global scope measurement accuracy, the data amount of spatial analysis application data sources is also gradually increase. Although in the past few decades, the performance of the CPU gradually improved and floating-point computing power achieved a higher level by continuously improve the CPU production process, the attendant problem of heat dissipation and power consumption impede the significantly increase of CPU clock frequency, limit the perform ability of single CPU and slow the growth of CPU floating-point computing power. With respect to the ever growing spatial data, the slowly upgrading CPU floating point computing power is obviously inadequate, which affect the computing speed of spatial analysis and limit the application of many excellent spatial analysis operator.Facing the limitation on floating-point computing power of existing computing platforms and huge computing needs of each application domain, people began to explore other solutions, which lead the microprocessor been into the multicore era. The importance of parallel programming has become increasingly prominent and scientific researcher and developers in various fields have started to try to use parallel programming to speed up the calculation. Heterogeneous Computing is a kind of special form of parallel computing, the basic idea of which is to reduce the time needed to complete the computing tasks by connecting computing devices which have different function or performance via a high-speed network. With low cost, low energy consumption, strong scalability and some other features, heterogeneous computing are more suitable than the traditional homogeneous parallel computing for space analysis which has huge amounts of data to calculate. CPU+GPU heterogeneous computing platforms is the mainstream of heterogeneous computing platforms, which occupy the dominance of heterogeneous computing architectures in TOP500supercomputer list.At present, in addition to the lack of floating-point computing power, the difficulty of further development of spatial analysis is the calculation short of universality, accuracy and standardization. The map algebra is a powerful tool for spatial analysis because of its wide and profound mathematical foundations of algebraic, and it use algebra view the theory and method of processing and visualization of geographic information’s nature and process. As a kind of theory and method solving geographic information visualization of graphic symbols and spatial analysis by using transform and computing of raster image, map algebra can be better adapt to the dynamic analysis process of multi-dimensional or multi-source spatial data in global wide range environment.In this article parallel acceleration strategy of map algebra operators whose process is relatively fixed and data source is raster image which has intrinsic parallelism will be discussed based on CPU+GPU heterogeneous computing platforms from the space analyzing point of view, such as data partitioning, overlapped data transmission and computing and raster preprocessing. The detail research contents are as follows:(1) Using CUDA technology to parallelize the traditional implementation based on CPU of map algebra operators whose process is relatively fixed and data source is raster image which has intrinsic parallelism, and release the CPU computing ability by processing intensive floating-point arithmetic operations on the GPU which is more suitable to do this than CPU.(2) Select the appropriate data partitioning strategy to split the large amount of data raster image based on the calculation characteristics of operators to hide data transmission time by overlapped data transmission and computing. Optimizing the data partitioning strategy to achieve the best hidden effect.(3) Design a suitable raster data storage structure for map algebra operators by making a research of data transmission between memory and video memory, which can read raster data by block rather than pixel, and exclude the data use to help to display efficiently in traditional raster data storage structure.At last, to verify the parallel acceleration strategies in this paper, a series of experiments have been performed, including data partitioning, overlapped data transmission and computing.