Parallelization and Performance Optimization of Radar Extrapolation Algorithm with OpenCL

Abstract

Extrapolation based on radar data is among the principal methods of weather now-casting, whereas the availability in business is severely affected by its high computational complexity and poor real-time. In order to address the problem, this paper presents the bottlenecks of the traditional algorithm process, and discusses how to parallel the radar extrapolation algorithm. It also illustrates some steps for optimizing the parallelized algorithm depending on different hardware environments and properties of OpenCL framework to achieve further improvement of calculating performance. Some methods such as optimizing the relationship between work group size and memory access performance, using local memory as high speed caches, hidden CPU execution time and switches branch structure were used to significantly improve the efficiency of the extrapolation process. The results show that by the series of optimization, the computational performance is boosted by more than 15 ~ 22 times under the situation of equivalent power consumption.