ML-based Adaptive Prefetching and Data Placement for US HEP Systems

School of Computing (SoC), University of Nebraska-Lincoln (UNL)

^** e-mail: saisuman1@acm.org
^*** e-mail: sbarla2@huskers.unl.edu
^**** e-mail: ramamurthy@unl.edu
^† e-mail: dweitzel@unl.edu

Published online: 7 October 2025

Abstract

Although benefits from caching in US HEP are well-known, current caching strategies are not adaptive i.e they do not adapt to changing cache access patterns. Newer developments such as the High-Luminosity - Large Hadron Collider (HL-LHC), Deep Underground Neutrino Experiment (DUNE), a steady move toward streaming readout based Data Acquisition systems (DAQs) will increase the data production exponentially and hence burden the storage, compute & network infrastructures. Moreover, existing caching frameworks are optimized to reduce latency, but not optimized for storage. This, in combination with limited cache capacities relative to total data, makes it difficult to achieve data locality.

In this work, we present Machine Learning-aided (ML) caching strategies. Specifically, we first present a Long Short-Term Memory-based (LSTM)hourly and multi-step cache usage prediction. Second, we present an hourly file-level access prediction model based on CatboostRegressor. To date, most ML-based cache prediction strategies in HEP have focused on daily cache usage and limited works tackled hourly cache usage and even fewer strategies addressed hourly file-level access prediction. File-level access prediction allows for the design of intelligent prefetching and data placement strategies with fine-grained control. We validated our cache prediction strategies using data collected from SoCal MINI caches in August 2024. We are currently extending the WRENCH simulator to reflect the US HEP ecosystem at the storage, network and compute levels. We plan to deploy our cache prediction strategies into WRENCH and later perform extensive analysis with complex data access patterns and candidate infrastructure configurations.