Workload-Aware I/O Optimization in Distributed Object Storage Systems

Abstract

High-performance computing (HPC) applications are increasingly dominated by data-intensive workloads that place new demands on storage subsystems. These workloads frequently diverge from traditional assumptions about I/O behavior, exhibiting irregular, small, and read-heavy access patterns that challenge the performance of distributed file systems. This dissertation explores empirical and predictive approaches to optimizing I/O performance in distributed object storage systems, focusing on workload-aware configuration strategies. This work systematically examines bandwidth and latency variations in the Ceph File System, built on top of the Ceph Distributed Object Store, due to diverse I/O patterns towards formalizing a data-driven relationship between latency, bandwidth, I/O size, and block-to-file size ratio. Our findings reveal that variations in system and application-level parameters can lead to substantial performance shifts, motivating the need for automated configuration support. We train a model for predicting bandwidth and estimating latency in similar workloads, providing storage administrators and application developers with actionable tuning recommendations, offering valuable insights into optimizing HPC storage I/O. Next, this dissertation focuses on the file striping strategy in CephFS, which utilizes a Raid-0 like policy to stripe incoming data over object(s) which are then mapped to the OSD(s). Currently, this striping strategy is determined at the time of cluster deployment and remains unchanged. We assess the impact of different stripe configurations by varying the object size, stripe count and stripe unit, on I/O performance across diverse workload types and access patterns. Our results highlight performance regimes where careful tuning yields significant gains, and expose inefficiencies in static, one-size-fits-all striping strategies. Overall, the focus of this dissertation is on the practical tuning of object storage systems grounded in empirical and predictive techniques enabling more responsive and efficient I/O performance in data-centric HPC workloads.