A Multi-label Classification Approach to Localization of Multiple Node Failures in Local DC Networks

The wide adoption of network based IT services to support operations and services have driven organizations to deploy local data center (DC) infrastructure and networks. Monitoring the proper functioning of such networks is of critical importance, specially in the event of failures. Timely detection and localization of the failed devices shorten the repair times and guarantee normal operation of infrastructure and services. In this work we propose a data-driven multiple failure localization approach based on device features obtained through passive monitoring. Namely, we set the localization problem as one of multi-label classification using high dimensional and high resolution data that is increasingly available with modern devices. Our results show that using simple base classifiers, the proposed methodology can yield high Hamming accuracy and acceptable compromise on false alarms, without relying on active monitoring.