On-line SVM traffic classification.
ABSTRACT A wide range of traffic classification approaches has been proposed in the last few years by the scientific community. However, the development of complete classification architectures that work directly in real-time on high capacity links is limited. In this paper we present the implementation of a machine-learning technique (SVM), one of the most accurate but most computationally expensive mechanisms, on the CoMo project infrastructure. We show the computational time required to process different traffic traces and the optimization steps we adopted to improve the performance of the system and achieve real-time classification on high-speed links.
- SourceAvailable from: S. Vaton[Show abstract] [Hide abstract]
ABSTRACT: Analyzing the composition of Internet traffic has many applications nowadays, like tracking bandwidth‐consuming applications, QoS‐based traffic engineering and lawful interception of illegal traffic. Even though many flow‐based classification methods, such as support vector machines (SVM), have demonstrated their accuracy, few practical implementations of lightweight classifiers exist. We consider in this paper the design of a real‐time SVM traffic classifier at hundreds of Gb/s to allow online detection of categories of applications. We also implement a high‐speed flow reconstruction algorithm able to handle one million concurrent flows. The solution is based on the massive parallelism and low‐level network interface access of FPGA boards. We find maximum supported bit rates up to 408 Gb/s for classification and up to 20 GB/s for flow reconstruction for the most challenging trace. Results are confirmed using a commercial Combov2 board with a Virtex 5 FPGA. Copyright © 2014 John Wiley & Sons, Ltd. We design and implement a lightweight traffic classifier based on Support Vector Machines (SVM) using the hardware acceleration of an FPGA to support data rates of hundreds of Gb/s. To do so, we develop a custom high‐speed flow storage mechanism able to support one million simultaneous flows using a limited memory. Results are confirmed using a commercial FPGA board.International Journal of Network Management 07/2014; 24(4). · 0.51 Impact Factor
- [Show abstract] [Hide abstract]
ABSTRACT: Analyzing the composition of Internet traffic has many applications nowadays, like tracking bandwidth-consuming applications or QoS-based traffic engineering. Even though many classification methods, such as Support Vector Machines (SVMs) have demonstrated their accuracy, the ever-increasing data rates encountered in networks are higher than existing implementations can support. As SVM has been proven to provide a high level of accuracy, and is challenging to implement at high speeds, we consider in this paper the design of a real-time SVM traffic classifier at hundreds of Gb/s to allow online detection of categories of applications. We show the limits of software implementation and offer a solution based on the massive parallelism and low-level network interface access of FPGA boards. We also improve this solution by testing algorithmic changes that dramatically simplify hardware implementation. We then find theoretical supported bit rates up to 473 Gb/s for the most challenging trace on a Virtex 5 FPGA, and confirm them through experimental performance results on a Combov2 board with a 10 Gb/s interface.IEEE Transactions on Network and Service Management 09/2014; 11(3):278-291.
- [Show abstract] [Hide abstract]
ABSTRACT: Traffic classification plays an important role in traffic management. To traditional methods, P2P and encryption traffic may become a problem. Support Vector Machine (SVM) is a useful classification tool which is able to overcome the traditional bottleneck. The main disadvantage of SVM algorithms is that it's time-consuming to train large data set because of the quadratic programming (QP) problem. However, the useful support vectors are only a small part of the whole data. If we can discard the useless vectors before training, we are able to save time and keep accuracy. In this article, we discussed the feasibility to remove the useless vectors through a sequential method to accelerate training speed when dealing with large scale data.Journal of the Korea Society of Computer and Information. 01/2012; 17(3).