Fast Clustering of Radar Reflectivity Data on GPUs

In short-term weather analysis, we use clustering algorithm as a fundamental tool to analyze and display the radar reflectivity data. Different from ordinary parallel k-means clustering algorithms using compute unified device architecture, in our clustering of radar reflectivity data, we face the dataset of large scale and the high dimension of texture feature vector we used as clustering space. Therefore, the memory access latency becomes a new bottleneck in our application of short-term weather analysis which requests real time. We propose a novel parallel k-means method on graphics processing units which utilizes on-chip registers and shared memory to cut the dependency of off-chip global memory. The experimental results show that our optimization approach can achieve 40× performance improvement compared to the serial code. It sharply reduces the algorithm’s running time and makes it satisfy the request of real time in applications of short-term weather analysis.