Análise de desempenho sequencial de sistemas que envelhecem e rejuvenescem

Abstract

Sequential performance analysis aims to evaluate performance indicators online. The process interrupts its execution according to a predefined stop rule, as soon as the number of observed anomalies reaches the stop rule, which should produce an alarm. Traditional sequential performance analysis techniques include CUSUM and sequential probability ratio test (SPRT). More recent techniques include the bucket algorithm, where tokens are accumulated in buckets when the system degrades and removed when the system naturally recovers. If the number of tokens in the system reaches a limit, an alarm will be triggered. The objective of this work is to analyze the sequential performance analysis algorithm (bucket algorithm) applied to systems that age and rejuvenate in light of three approaches: 1) Without system information, random rejuvenation, when rejuvenation occurs at any time and without direct relation to the system state, 2) Partial system information, when we implement a Markov chain model with three dimensions (anomaly source state, anomaly detection algorithm, system queue state), in this approach rejuvenation is signaled to the system by the bucket overflow, representing system criticality and the need for administrator intervention, and 3) Complete system information, where, using approach 2, simulating a hypothetical situation of knowledge of the anomaly source state and system queue state, which is reflected by the bucket state, in this approach, once in the anomalous state, rejuvenation is applied. Thus, approaches 2 and 3 aim to simulate a closed-loop control, where system output values that deviate from the expected generate a correction action, aiming at normalizing the system operation.