The dCache storage system at Brookhaven National Laboratory (BNL) serves as a critical cache for the ATLAS collaboration, enabling efficient access to petabytes of data located on tape, remote repositories, and cold storage. Effective cache management is vital to minimize access latency, particularly as operators have observed persistent high-demand datasets that warrant prolonged retention (“pinning”) in disk cache. This study evaluates machine learning (ML) techniques to automate dataset pinning decisions by predicting future access patterns. Our models, which integrate temporal trends and request-specific features, achieve predictive errors significantly below the inherent variability of dataset access patterns. We further explore dynamic updates to these predictions using real-time dCache access logs, enabling adaptive pinning strategies for high-priority datasets. Ongoing work focuses on validating system-wide performance gains under realistic user workloads, with the goal of optimizing resource utilization for large-scale scientific data infrastructures.
Aldrich et al. (Tue,) studied this question.