Exploiting a Self-learning Predictor for Session-based Remote Management Systems in a Large-scale Environment

Abstract

Session-based remote management systems, e.g., customer premises equipment (CPE) WAN management protocol (CWMP), have predictable task counts in a session and each CPE only accesses its own data. When the systems are used in large-scale environments, a static load balancing (LB) policy can be applied with fewer session migrations. Nevertheless, unexpected crash events, e.g., software bugs or improper management, would cause the LB policy to be reassigned so as to degrade the system performance. A self-learning predictor (SLP) is thus proposed in this work to predict unexpected crash events and to achieve a better system performance in terms of resource usage and throughput. Specifically, the SLP records and monitors all crash patterns to evaluate the system stability. Moreover, the relation flags and probabilities of all crash patterns are dynamically updated for quick comparisons. If the SLP finds the current pattern is similar to a crash pattern, a migration request is raised to the load balancer to prevent performance degradation caused by the incoming crash. The simulation results indicate that a better system performance is obtained in a large-scale environment with the proposed SLP, especially as the number of servers in each cluster node increases.