Large-Scale Inference of Network-Service Disruption upon Natural Disasters

Large-scale natural disasters cause external disturbances to networking infrastructure that lead to large-scale network-service disruption. To understand the impact of natural disasters to networks, it is important to localize and analyze network-service disruption after natural disasters occur. This work studies an inference of network-service disruption caused by the real natural disaster, Hurricane Katrina. We perform inference using large-scale Internet measurements and human inputs. We use clustering and feature extraction to reduce data dimensionality of sensory measurements and apply semi-supervised learning to jointly use sensory measurements and human inputs for inference. Our inference shows that after Katrina, approximately 26% of subnets were inferred as unreachable. We find that 57% of unreachable subnets were small subnets at the edges of networks, and 45% of these unreachabilities occurred after the landfall. The majority (73%) of unreachable subnets lasted longer than four weeks showing that Katrina caused extreme damage on networks and a slow recovery. Network-service disruption is inevitable after large-scale natural disasters occur. Thus, it is crucial to have effective inference techniques for more understanding of network responses and vulnerabilities to natural disasters.