Abstract: As a classic algorithm in data mining, the clustering algorithm is often adopted in analysis of radar reflectivity data. However, it is time-consuming while facing dataset of large scale and high dimension. Recently, several studies have been conducted to make effort in parallclization or optimization of the clustering algorithm on GPUs. Although these studies have shown promising results, one important factor`program inputs-in the optimization is ignored in optimization. We took the program inputs in consider as a factor for optimization of the clustering algorithm on GPUs. By observing the distribution feature of the input radar reflectivity data, we found that the ability to adapt to inputs is important for our application to achieve the best performance on GPUs. The results shows that our approach can gain a 20%一40% performance increment, compared to previous parallel code on GPUs, which makes it satisfies the requirement of real-time application well.

Key words: Clustering algorithm, Real-time, Input aware, GPU, CUDA