A method of multidimensional software aging prediction based on ensemble learning: A case of Android OS

Context: Software aging refers to the phenomenon of performance degradation, increasing failure rate, or system crash due to resource consumption and error accumulation in software systems running for a long time. It has become the key factor affecting software systems’ sustainability. Due to its complex formation reasons, precisely predicting the aging state in actual execution is hard but crucial for enabling proactive measures before a catastrophic situation. Machine learning (ML) has been employed on this issue. Objective: However, previous ML-based prediction methods are single-threaded in the whole process, posing challenges in delivering the desired performance facing diverse user scenarios. To alleviate this problem, we propose a multidimensional software aging prediction method based on ensemble learning (MSAP). Method: In the framework of MSAP, five dimensions, including datasets, labeling metrics, labeling thresholds, algorithms, and model decisions, are extracted and diversified according to aging characteristics and application situations. Results: Plenty of experiments have been conducted on Android devices from three distinct vendors. When subjected to identical workloads, MSAP demonstrates comparable performance to most unidimensional models. While under varied workloads, MSAP outperforms unidimensional models whose performance drops dramatically, demonstrating enhanced adaptability and predictive accuracy. Conclusion: MSAP shows exceptional stability while concurrently upholding outstanding prediction precision across a spectrum of user scenarios. It has better generalization characteristics and application prospects.