Design And Implementation Of The Network Traffic Feature Extraction Based On GPU

Network traffic classification is playing an increasingly important role in terms of enhancing the network management and controllability. With the development of network applications, real-time and more accurate traffic classification technologies are needed. Recently, researchers in this field try to use the technology of machine learning to solve this problem, and achieve great results. But as one of the important part of the machine learning, the feature extraction requires high computational complexity and time-consuming; this becomes the bottleneck of the machine learning when dealing with the real-time traffic classification.In recent years, the rapidly development of the GPU hardware architecture, makes the capacity of floating-point arithmetic and the parallel computing of GPU much better than CPU. So the GPU is widely used for accelerating the large-scale parallel processing and scientific computing. In particular, the CUDA programming model introduced by the NVIDIA, provides large amount of API functions, which makes the parallel processing of the GPU works better.In this paper, we first analyze the architecture of the GPU and the parallel programming model. Then, we introduced the serial process of feature extraction algorithm. By compared the scale and the characteristics of each calculation task, we analyzed feasibility of parallelism of the serial process. Based on that, we designed a parallel algorithm according to the CUDA, and optimized the algorithm by the streaming technology and the features of the GPU heterogeneous execution. In the last, we deployed our feature extraction system on Linux platform and compared our parallel algorithm and optimized version with traditional serial processing algorithm in terms of time-consuming. Experimental results show that, when dealing with large amounts of flow records, optimized parallel algorithm can speed up more than2times over the sequential approach in the overall algorithm execution time, which gained a better performance advantage.