Workshops Archive

Network-Aware Caching in Distributed CMS XRootD Systems via Data-Driven Forecasting


Workshop: 13th Annual International Workshop on Innovating the Network for Data Intensive Science — INDIS

Authors: Sarat Sasank Barla, Byrav Ramamurthy, and Derek Weitzel (University of Nebraska-Lincoln)

Abstract: High-Energy Physics experiments such as the Compact Muon Solenoid distribute petabyte-scale datasets across geographically dispersed sites, imposing sustained load on wide-area research networks. Although in-network caching mitigates repeated transfers, deployed policies remain reactive rather than anticipatory. In this work, we propose a two-stage predictive caching framework trained on 8.4 million XRootD session records from 25 cache nodes across four domains. In Stage-1, a gated recurrent unit model forecasts near-future dataset accesses. In Stage-2, directories and files within predicted datasets are identified through non-parametric resolution. Simulation results show that an architecturally decoupled side-cache preserves baseline performance. Over 23 days at a single cache node, it displaces 115 GB of wide-area traffic, reducing wide-area volume by 1.6% and improving the hit rate by 1.4–1.6 percentage points. This study demonstrates that predictive caching depends not only on prediction accuracy, but also on selecting prediction granularity and integrating predictions with cache operation.


Back to 13th Annual International Workshop on Innovating the Network for Data Intensive Science — INDIS Archive Listing

Back to Full Workshop Archive Listing