Adaptive polling intervals for network monitoring

Abstract

Monitoring is an essential task for the management of any network and several network functions dependent on it. In Software Defined Networking (SDN) is no different, even with all the benefits provided by it. Network monitoring helps to understand patterns in network traffic, allowing to capture its state and enhance its configuration. Furthermore, it aids the infrastructure management, the discovery of bottlenecks, the location of problems related to software and hardware, and enforces SLAs (Service Level Agreement). Cloud services are spread around the globe and are used constantly by several companies. However, ensure the safety and quality of these services is a difficult and non-stop chore. Network monitoring has a major part in assuring that the services are executed without problem. Ideally, it is expected that the monitoring solution does not overload the network, scales together with the network causing minimal impact, has a controlled resource usage and works for every network device independent of the vendor. The vendor lock-in problem was solved by the OpenFlow protocol, the main and default SDN protocol. However, the others aforementioned problems are dependent on the monitoring solution implementation and used strategies. Furthermore, to the best of our knowledge, no SDN monitoring solution solved these problems ensuring all the needed network monitoring capabilities. In order to address this gap, we proposed SHAMon, a network monitoring solution for Software Defined Networking, created to provide fine-grained and precise information, generating minimal network and resources overhead. We propose the use of the binomial algorithms used for congestion control by the TCP protocol to control our statistics request mechanism. We use all the OpenFlow features to aid in getting timely and refined statistics. Additionally, we use a proxy-server architecture to allow our solution to scale with the network infrastructure. To validate the use of binomial algorithms, we defined a set of experiments. The first set of experiments was intended on analyze two functions of our solution, the former being how variations of parameters of the congestion control algorithms are related to how much the polling time vary. The latter was how variations of the thresholds that limit the execution of the congestion control algorithms are related to the decision to vary the polling time. In addition, we did a comparison experiment with three other solutions, Payless, periodic polling, and OpenNetMon in the same network scenario. The major objective of this experiment was to highlight some of our qualities. The results from both experiments were good, the former showed us the behavior of the binomial algorithms and our solution. Our solution had a low error in network measurement, varying from 26% to 15%. The latter showed that our monitoring overhead was three times lower than the others, validating that our solution is both accurate and not network consuming.