Learning-Based Cache Allocation across Shared HPC and AI Workloads: A Reproducibility-Centred Evaluation

Abstract

Scientific simulations and AI training can compete for the same storage cache even when aggregate throughput appears satisfactory. The difficulty is particularly visible when a system is transferred to a new setting. The objective is to assess whether learning-based cache admission improves workload fairness without adding excessive overhead. Relevant methodological and data-quality literature informs the design of the sampling and validation strategy [3–7]. The proposed method is a controlled trace-replay benchmark combining training-data reads with parallel simulation checkpoint traffic. Comparators will include static partitions, recency-based eviction and learned admission/prefetching policies. Related enterprise-storage studies inform the assessment of operational reliability [1] and the deployment assumptions adopted for system-level comparison [2]. The reproducibility extension records dataset versions, model settings, exclusion rules and provenance before any held-out evaluation. Statistical and computational sources help define the reproducibility checks and comparison procedures [8–12]. The primary outcomes are job-level tail latency, backend I/O volume, useful cache occupancy and fairness across applications. Potential biases include working-set changes, cold starts, prediction CPU cost and network contention. An independent replication exercise would repeat the main analysis from the documented artefacts and quantify disagreements. Additional research informs the prespecified stress tests and limits on interpretation [13–17]. The intended contribution is an auditable research protocol rather than an unsupported claim of effectiveness; related work frames its transferability limits [18–23].