Local acceleration in Distributed Geographic Information Processing with CUDA

DGIP (Distributed Geographic Information Processing) has become a new tendency of GIS (Geographic Information System) recently. DGIP focuses on how to organize and process a series of geographic resources in distributed computing environment and now existing research is mainly carried out from a global point of view. But it is noticeable that each computing node in distributed computing environment will carry a heavy load with growth of data quantity. So this paper concentrates on how to make each computing node fulfill the subtask more quickly to achieve efficient local acceleration. The paper designs a prototype for distributed remote sensing image processing and achieves local acceleration in each computing node with CUDA (Compute Unified Device Architecture). Firstly, the paper introduces the distributed procedure of the prototype and overviews the architecture and programming model of CUDA. Then the paper takes Mean Filter as an example to design and implement the parallel program with CUDA to accelerate the procedure of remote sensing image processing in each node. To evaluate the performance of the local acceleration, the paper carries out a group of comparative tests between the parallel implementation with CUDA and the conventional implementation. The results demonstrate that the local acceleration with CUDA runs more than 20 times faster than conventional process.