Alarm Mechanism for Anticipated Detection of Network Unavailability in IP Networks Through Time Series Analysis

With organizations and individuals increasingly depending on the Internet, failures in subnetworks may affect important services such as for economic, health, educational and governmental purposes. Also, as the complexity of the Internet increases, efficient monitoring and automatic preventive measures play vital roles in avoiding network services interruption. In this work, we propose an alarm mechanism based on time series analysis, which monitors entire IP networks with low overhead and anticipates strong degradation of network performance. When applied on 9 operational Internet Protocol (IP) networks with nodes spread worldwide, our mechanism was able to anticipate cases of 100% loss 15 minutes earlier with more than 99% of Accuracy and less than 0.05% of FalsePositive Rate in the vast majority of the networks we tested with months of monitoring. In the worst case scenario we found, with Approximate Entropy (ApEn) indicating severely random behavior of loss measurements, our mechanism reached 96.5% of Accuracy and 2% of False-Positive Rate. Based on such encouraging results, we believe our alarm mechanism will support daily operations in IP networks, resulting in enhanced resiliency and enabling more intelligent service provisioning.