Geographically replicating objects across multiple data centers improves the performance and reliability of cloud storage systems. Periodic replication is a common method to reduce synchronization cost and improve system throughput. Existing periodic replication methods are generally static and lack quantitative analysis models, leading to sub-optimal synchronization costs.
To solve the problems, a research team led by Limin Xiao and Liang Wang published their new research on 15 October 2024 in Frontiers of Computer Science co-published by Higher Education Press and Springer Nature.
The team proposed a dynamic periodic synchronization strategy by establishing a quantitative analysis model. The model and synchronization strategy were verified and tested through experiments in the Ali Cloud environment, which demonstrated their effectiveness. Compared to existing static methods, the synchronization strategy can adapt to different workloads and greatly reduce synchronization costs.
In the research, they aim to reduce the synchronization cost by quantifying and analyzing the tradeoff between consistency and synchronization cost. To achieve this, they first introduced the consistency quantification analysis model, TTK, which explores the average staleness of the data retrieved by a read request under different synchronization cycles and workloads using the T-Freshness, T-Visibility, and K-Staleness metrics. Then, the team developed a synchronization cost model, CMC, which models the tradeoff between consistency and synchronization cost, and derives the optimal synchronization period that minimizes the synchronization cost. Based on these models, they proposed Sync-Opt, a dynamic synchronization mechanism that adapts the synchronization period to different workloads, and achieves the best tradeoff between consistency and synchronization cost.
The experiments were carried out in Ali Cloud environment. The experimental results show the validity of the quantitative analysis model, and the dynamic synchronization strategy can adapt to different workloads and greatly reduce the synchronization cost.
Future work can focus on improving synchronization performance, that is, theoretically or quantitatively analyze their dynamic synchronization mechanism in terms of latency or other performance metrics.
DOI: 10.1007/s11704-023-2625-8
