Combining Learned and Highly-Reactive Management

Learned models of behavior have the disadvantage that they must be retrained after any change in system configuration. Autonomic management methods based upon learned models lose effectiveness during the retraining period. We propose a hybrid approach to autonomic resource management that combines management based upon learned models with “highly-reactive” management that does not depend upon learning, history, or complete information. Whenever re-training is necessary, a highly-reactive algorithm serves as a fallback management strategy. This approach mitigates the risks involved in using learned models in the presence of unpredictable effects, including unplanned configuration changes and hidden influences upon performance not considered in the learned model. We use simulation to demonstrate the utility of the hybrid approach in mitigating pitfalls of both learning-based and highly-reactive approaches.