Compressed incremental checkpointing for efficient replicated key-value stores

The prominent cloud services rely on geographically distributed nodes running replication and other fault-tolerance mechanisms so as to provide flawless availability and dependability. In this paper, we address the communication cost of the well known primary-backup replication protocol, and propose compressed periodic incremental checkpoint algorithms to achieve improved throughput. We set up a replicated key-value store on geographically distributed nodes of the PlanetLab platform, and developed compressed incremental checkpointing algorithms to support primary-backup replication. By considering performance metrics of interest including blocking time, checkpointing time, compression ratio, compression/ decompression times, we conducted a comprehensive analysis. We used the well-known benchmarking tool YCSB and established different sample workloads to test where each workload represents diverse plots. Our findings indicate that Zstd is the most competent compression method under all scenarios and through comparing with an uncompressed approach we point out that compressing the communication data disseminated from the primary replica coupled with the periodic incremental checkpointing algorithm not only decreases the average blocking time up to 5% but it also improves the overall system throughput by 4% compared to the no compression case.