Olsync: Object-level tiering and coordination in tiered storage systems based on software-defined network

Abstract

With the adoption of new storage technologies like NVMs, tiered storage has gained popularity in large-scale, hyper-converged clusters. The storage back-end of hyper-converged systems supports data storage on devices such as SSDs and HDDs, yet lacks fine-grained tiered storage solutions. For example, Ceph selects storage nodes based primarily on limited criteria, such as node storage capacity, disregarding the diverse performance characteristics of various storage media. In this study, we introduce Olsync, an object-level tiering and coordination system designed to enhance storage resource utilization and data access performance. Specifically, Olsync employs PIPO (Packet-In-Packet-Out), an innovative network communication framework based on Software-defined Networking (SDN), to collaboratively optimize both the network control plane and underlying data plane. Additionally, Olsync can offer efficient object-level tiering and coordination services using the global views obtained by PIPO (e.g., data access patterns and interfering object requests) to make tiered storage and performance optimization decisions. We incorporated the Olsync prototype into Ceph and performed a thorough comparison with contemporary state-of-the-art systems. The evaluation results demonstrate that Olsync significantly enhances system response time (up to 68%), I/O throughput (up to 24%), and 99th percentile latency (up to 16%) in various environments.