Application-Aware I/O Prediction and Optimization for HPC and AI Workloads
Files
TR Number
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
High-performance computing (HPC) systems are essential for executing large-scale scien- tific and artificial intelligence workloads that demand extensive parallelism across compute, storage, and memory resources. Modern systems, such as Aurora at Argonne National Lab- oratory and the exascale Frontier system at Oak Ridge National Laboratory, have significantly expanded computational and I/O capabilities. However, storage and memory subsystems often struggle to keep pace with the scale, diversity, and dynamic behavior of modern applications. This gap creates performance bottlenecks that are difficult to diag- nose and mitigate using rigid, application-agnostic I/O management strategies. Addressing these challenges requires a deeper understanding of how application semantics, execution configurations, and workload behavior interact with storage and memory systems. This thesis addresses these challenges across three complementary directions. First, we model application I/O behavior using historical execution patterns and user-specified parameters. Through analysis of E3SM and LAMMPS workloads, we show that prior application con- figurations are strongly correlated with future I/O behavior, enabling accurate prediction over 1–10-day windows with uncertainty below 0.25 and 80–90% prediction accuracy. Sec- ond, we introduce IO-Sense, an agentic I/O reasoning pipeline that combines Darshan logs with application-level configuration semantics to summarize large-scale traces and explain how configuration knobs influence I/O patterns. On AMReX-based HPC applications, IO- Sense achieves over 90% log similarity with the ground-truth log derived from the knob change for most evaluated models and demonstrates that finer-grained I/O counters can improve reasoning accuracy, though at the cost of higher inference overhead. Third, we study I/O and memory behavior in deep learning training workloads through SHADE and SODA-LLM. SHADE introduces a rank-based importance sampling policy that improves data locality and cache efficiency for large-scale computer vision training workloads. Build- ing on this direction, SODA-LLM targets transformer-based language model training/fine- tuning by exploiting the spatial and temporal stability of activation importance to guide token-level activation retention, offloading, and prefetching during the backward pass. The proposed system reduces end-to-end training latency by up to 20% while lowering memory usage by 5%. Together, these projects show that application-aware reasoning can improve prediction, diagnosis, and optimization across HPC and AI workloads. By connecting appli- cation configurations and execution semantics to observed I/O and memory behavior, this thesis provides a foundation for more adaptive storage and memory management strategies in next-generation computing systems.