A Dynamic Polling Strategy Based on Prediction Model for Large-Scale Network Monitoring

The scale of modern networks grows exponentially, challenging our ability of efficiently managing large-scale network systems. Building an efficient monitoring system for such a large-scale network to report problems is difficult because of the scale of the network. Conventional approaches based on a fixed period polling strategy cannot fast adapt to the change of network and fail to discover abnormal nodes quickly. To tackle this problem, this paper proposes a dynamic polling strategy. Our model is based on wavelet packet decomposition and support vector regression. It uses the wavelet packet decomposition techniques to decompose the comprehensive load series into several subseries whose rule is relatively easy to be learned. Next, it uses the support vector regression method to predict these sub-series. It then dynamically adjusts the polling frequency through accurate forecasting of the network state. The proposed approach has been applied in a live, large-scale commercial network consisting of more than six thousand nodes. Experimental results show that this algorithm maintains a high sensitivity of monitoring and greatly improve the accuracy of performance data.