Anticipating minimum resources needed to avoid service disruption of emergency support systems

Abstract

Advancement in the optimization of resources motivates us to study new mechanisms for the automated and elastic adaptation of virtual computer and network systems. Thus we designed the Autonomic Resource Control Architecture (ARCA), which considers the workload of the controlled system together with events notified by external detectors to perform its work. However, there is a delay between the occurrence of an event and the adaptation of the system. In this paper we propose a mechanism to enable ARCA to anticipate the minimum resource amount required by the controlled system under different situations by using a Machine Learning (ML) mechanism. Related solutions only consider the monitoring data provided by the controlled system, require a long learning period, are fragile to topology changes, and are unfeasible for real time operations. We propose to resolve such problems by using a threshold-based method to self-assess and self-correct the knowledge of our ML-based method, thus achieving self-learning qualities and ensuring that correct decisions are issued. Moreover, we set computational boundaries to the algorithm, so it runs within acceptable performance limits. Finally, we demonstrate its qualities by executing a simulation on a generated dataset following a demonstrated behavior, showing that the anticipation method results in no drop of client requests, using just 15% more resources than a threshold-based method.