GPU-oriented stream data mining traffic classification

Abstract

Network Management depends on precise characterization of the traffic profile of networked applications. When the identification and classification of network flows is done using machine learning, the characterization of traffic still requires an approach that is capable of providing a balance between accuracy and processing speed in real-time scenarios. This paper proposes an architecture to classify network traffic based on Stream Data Mining techniques using Graphic Processing Units (GPU), in order to meet the requirements of both classification accuracy and speed. Our proposal combines the characteristics of data mining techniques with a continuous stream of input data, and with high processing performance GPU architecture. Results show that our approach provides accuracy comparable to or better than existing related work (e.g., above 95%) while ramping up performance (e.g., up to 62x speed up), comparing the different implementations of our approach. These facts allow the deployment of the proposed technique to the realtime management of high speed backbone links. I.INTRODUCTION he diversity of applications for computer networks has increased considerably over the last decade. Additionally, the management and monitoring of networks depends on the precise characterization of the traffic profile generated by the network applications [1] [5]. To this end, there are three approaches that can be used to classify network traffic [5], namely port-based, packet payload inspection, and flow-based techniques. Port-based traffic identification has a low computational cost and is very inaccurate [11]. On the other hand, packet payload inspection presents very accurate results, but it requires high computational processing. Additionally, it cannot identify encrypted traffic and, in some countries, it cannot be deployed due to legal restrictions. Therefore, flowbased classification becomes the best alternative, since it presents both high accuracy and relatively low computational demands [20]. This type of traffic classification involves classifying the flow (the aggregation of network packets, defined by a 5-tuple [5]) according to its features.However, the use of flow-based techniques shows us other important issues in the design of a flow-based classifier. The capacity of the collector machine to store flow records and the computational performance of the flow classifier are relevant. For example, a router on a 10 Gbps link can deal with approximately 10 thousand flows simultaneously [3]. From a classification perspective, a high number of flows imply a huge computational cost. This is because the memory and performance requirements increase as the number of flows

GPU-Accelerated Support Vector Machines for Traffic Classification

“…Since GPU shows good performance in image processing and scientific computing, GPU is applied to speed up machine learning models [10]. The parallelization of machine learning models can be carried out from two perspectives: model parallelism and data parallelism [13].…”

“…Zhou et al stored the tree models through the shared memory of GPU and put forward C4.5 decision tree method based on GPU [5]. Lopes et al suggested a GPU-Oriented Stream Decision Tree (GSDT) and used SIMD architecture to parallel classify network traffic [13].…”

“…With the growth of Internet traffic from TB level to ZB level, traffic classification needs to speed up to 10000 per second [18]. Due to SVM, neural networks, decision tree and other machine learning models consume large CPU resources, it is difficult to solve a large-scale and real time network traffic classification problem on traditional commercial machine platform [13].…”