Context Matters: Constant-Time File Lifecycle Prediction from File Creation Call Stacks

Many I/O optimizations, such as data placement on hierarchical storage systems, depend on knowledge of file lifetimes and access patterns. However, such information is typically unavailable at file creation time or must be inferred through costly runtime monitoring. In this paper, we demonstrate that file creation context alone is a strong predictor of file lifetime and lifecycle behavior. We introduce CLiP, a Constant-time Lifecycle Prediction approach that enables the prediction of either a scalar file lifetime or a full file lifecycle distribution at the moment of file creation. Our method leverages the file creation function call stack hash as contextual information and performs lightweight prediction using a very small pre-trained per-application lookup table, incurring negligible runtime overhead compared to existing NN-based approaches. Moreover, in contrast to tracing-based techniques that always intercept all I/Os, once trained, CLiP only requires the instrumentation of file creation functions. Our evaluation on three representative HPC workloads (NAMD, Incompact3d and LAMMPS) shows that CLiP predicts ∼ 83.2% of lifetimes within ±20% of the observed values versus ∼ 82.0% for a pathname-based CNN baseline, while reducing prediction overhead by ∼ 99.7%, i.e., ≈ 292x faster. These results highlight that context at file creation time matters and that exploiting this context can open the door to practical and low-overhead I/O optimizations in large-scale systems.