Robust resource allocation for online network monitoring

Building robust network monitoring applications is hard given the unpredictable nature of network traffic and continuous growth of link speeds, data rates and complexity of traffic analysis tasks. Effective resource management techniques are now a basic requirement for this class of applications, which have to deal inevitably with the effects of extreme overload situations during their normal operation. In this paper, we present in detail the problems involved in the management of system resources in network monitoring and describe the design of a load shedding scheme that can efficiently handle extreme overload situations by gracefully degrading the accuracy of monitoring applications. Our method controls the resources allocated to each application by dynamically adjusting the sampling rate based on an online prediction model of the system resource requirements. We present experimental evidence of the robustness and performance of our system using real traffic traces and injecting synthetic traffic anomalies.