Learning-Based Cache Allocation across Shared HPC and AI Workloads

Abstract

Scientific simulations and AI training can compete for the same storage cache even when aggregate throughput appears satisfactory. This gap has consequences for interpreting performance outside carefully curated evaluation settings. This study aims to assess whether learning-based cache admission improves workload fairness without adding excessive overhead. The evaluation would follow a controlled trace-replay benchmark combining training-data reads with parallel simulation checkpoint traffic. The choice of evaluation units and exclusion criteria is informed by the broader methodological record [3–7]. The analysis will contrast static partitions, recency-based eviction and learned admission/prefetching policies. The design prioritises precise operational definitions and a straightforward baseline before introducing more complex alternatives. Prior work on storage operations provides technical context for validation [1], while complementary systems evidence motivates service-continuity checks [2]. Assessment will focus on job-level tail latency, backend I/O volume, useful cache occupancy and fairness across applications. The supporting evidence base motivates the chosen comparator conditions and robustness criteria [8–12]. Critical sensitivity checks address working-set changes, cold starts, prediction CPU cost and network contention. Further systems and modelling literature supplies context for the proposed sensitivity analyses [13–17]. Primary analyses would be reported by the main deployment or participant subgroup so that averaged results do not conceal systematic failures. The expected output is a reusable evaluation framework with transparent uncertainty and reporting rules, informed by wider reproducibility literature [18–23].