Abstract
Modern I/O applications that run on HPC infrastructures are increasingly becoming read and metadata intensive. However, having multiple applications submitting large amounts of metadata operations can easily saturate the shared parallel file system's metadata resources, leading to overall performance degradation and I/O unfairness. We present PADLL, an application and file system agnostic storage middleware that enables QoS control of data and metadata workflows in HPC storage systems. It adopts ideas from Software-Defined Storage, building data plane stages that mediate and rate limit POSIX requests submitted to the shared file system, and a control plane that holistically coordinates how all I/O workflows are handled. We demonstrate its performance and feasibility under multiple QoS policies using synthetic benchmarks, real-world applications, and traces collected from a production file system. Results show that PADLL can enforce complex storage QoS policies over concurrent metadata-aggressive jobs, ensuring fairness and prioritization.

Abstract

Abstract
Machine learning (ML) frameworks, such as PyTorch and TensorFlow, rely on data loaders to preprocess data before feeding it to accelerators. When preprocessing is inefficiently pipelined, GPUs can remain idle over long periods of time, leading to substantial training delays. For example, PyTorch’s default data loaders can cause up to 76% GPU idleness. A key bottleneck is the variability in preprocessing time across samples within the same dataset. Existing data loaders are oblivious to this variability, training all samples uniformly. In this case, a single slow sample can stall the entire batch, causing head-of-line blocking. We present MinatoLoader, a general-purpose data loader for PyTorch that accelerates training and improves GPU utilization under single-server, multi-GPU settings. It continuously prepares data in background and constructs batches by prioritizing fast-to-process samples, while slower samples are processed in parallel. Experiments conducted over NVIDIA V100 and A100 GPUs show that MinatoLoader accelerates training by up to 7.5× (3.6× on average) over PyTorch DataLoader and Pecan, and up to 3× (2.2× on average) over DALI. It also increases average GPU utilization from 46% with PyTorch to 90%, while preserving model accuracy and enabling faster convergence. © 2026 Copyright held by the owner/author(s)

Abstract

Abstract

Abstract