Convergence of HPC and Cloud Storage Systems for Deep Learning Workflows: An Evaluation of LustreFS and S3-Compatible Object Storage

High-performance computing (HPC) file systems and cloud-native object storage are essential for supporting large-scale deep learning (DL) workloads. While HPC file systems provide high-speed, low-latency access, cloud storage offers elastic scalability for heterogeneous I/O patterns. This paper introduces a converged HPC-cloud storage platform by integrating Lustre with S3-compatible object storage (MinIO). This integration is achieved through a lightweight, non-invasive gateway layer that requires no modifications to the underlying Lustre deployment. The architecture includes a lightweight gateway for interoperability and an adaptive hyperparameter optimization framework, based on a multi-armed bandit (MAB) approach, for dynamically tuning storage configurations.We evaluate the system using the MLPerf Storage benchmark suite, comparing it against Ceph and traditional XFS deployments, and analyze model-specific sensitivity to storage performance. Experimental validation demonstrates that our converged system achieves up to 33.3% higher read throughput and 40.7% lower metadata operation latency compared to Ceph in equivalent configurations.