Predicting z/OS Data Set Usage Patterns From SMF Records: A Machine Learning Feasibility Study

Abstract

This thesis explores whether machine learning can improve storage management in z/OS mainframe systems by predicting data set access patterns from SMF logs. Using a large dataset of over 5 million access events across 213,000 data sets, it finds that near-term access and access frequency can be predicted accurately, while time-to-next-access and behavioral clustering are ineffective in continuously active environments. The work highlights key methodological limits, such as scalability challenges and regression constraints, and shows that prediction reliability depends on the presence of idle periods, which are absent in dense systems. Overall, it delivers both a production-scale pipeline and a clear understanding of when data set usage prediction is feasible.