Self–Rejuvenation - an Effective Way to High Availability

Most computer users know that one of the most effective ways to recover from faults such as system crashes and performance degradation is simply to reset or reboot a computer. Although these methods might be expensive in terms of downtime, they are frequently most effective. Since commercial users in most cases cannot afford the restart, they tend to use periodic rejuvenation, which is restarting of processes, programs or entire systems in a preventive manner at the appropriate time for a given application. We propose self-rejuvenation, i.e., automatic rejuvenation based on failure prediction by means of modeling with Markov chains and Universal Basis Functions (UBF). Once the severity of a potential failure is determined by our models, the rejuvenation procedure may begin either by a process restart, a restart of application, reboot of the system, check-pointing, going back to recovery point or fail-over by reloading and recomputing an application on another computer. There is a strong correlation between failures and system dynamics. We show how system models can be constructed by analyzing error logs and ‘fever charts’ from system’s logs. Furthermore, we demonstrate the quality of our models based on real-system data. It will be shown that prediction-triggered self-rejuvenation is an effective way to reducing unavailability by an order of magnitude.