Enabling efficient, high-performance data access in parallel file systems (PFS) is critical for today's highperformance computing systems. PFS client-side I/O heavily impacts the final I/O performance delivered to individual applications and the entire system. Autotuning the key client-side I/O behaviors has been extensively studied and shows promising results. However, existing work has heavily relied on extensive number of global runtime metrics to monitor and accurate modeling of applications' I/O patterns. Such heavy overheads significantly limit the ability to enable fine-grained, dynamic tuning in practical systems. In this study, we propose DIAL (Decentralized I/O AutoTuning via Learned Client-side Local Metrics) which takes a drastically different approach. Instead of trying to extract the global I/O patterns of applications, DIAL takes a decentralized approach, treating each I/O client as an independent unit and tuning configurations using only its locally observable metrics. With the help of machine learning models, DIAL enables multiple tunable units to make independent but collective decisions, reacting to what is happening in the global storage systems in a timely manner and achieving better I/O performance globally for the application.

Modern high-performance computing (HPC) applications run exclusively on computational resources but share global storage systems. This design can lead to issues when applications use disproportional amount of storage resources compared to their allocated computing resources. An application running on a single compute node might consume excessive I/O bandwidth from a storage server, for example, by issuing numerous small, random writes. In doing so, it can hinder larger jobs that also write to the same storage server and are allocated many compute nodes, resulting in significant resource waste. A straightforward solution to prevent such an issue is to limit each application's I/O bandwidth on storage servers in proportion to its allocated computation resources. This approach has been implemented in existing parallel file systems using the Token Bucket Filter (TBF) mechanism. However, applying such methods in practice often results in lower overall I/O efficiency. HPC applications are known for generating short, bursty I/O requests. Therefore, strictly limiting I/O bandwidth proportionally either leads to wasted storage server bandwidth when applications are not performing I/O, or it prevents applications from temporarily utilizing higher bandwidth during bursty I/O phases. We argue that the goal of I/O control should be to maximize both the I/O performance of each application and the overall efficiency of the storage server, while ensuring fairness among jobs, such as preventing smaller jobs from blocking largescale ones. In this paper, we propose AdapTBF to achieve this objective. Building on the well-established TBF mechanism in modern parallel file systems (e.g., Lustre), AdapTBF introduces a decentralized bandwidth control approach using an adaptive borrowing and lending mechanism. We detail the algorithm, implement AdapTBF on the Lustre file system, and evaluate it using synthetic workloads modeled after real-world scenarios to demonstrate its effectiveness. Experimental results show that AdapTBF effectively manages I/O bandwidth for applications on storage servers while maintaining high overall storage utilization, even under extreme conditions.