Evaluation of a Tenant Level Checkpointing Technique for SaaS Applications

Tenant level checkpointing is a novel fault-tolerance technique proposed in our previous work for large-scale software-as-a-service applications. This paper evaluates the technique by a theoretical analysis of the technique and an empirical study using workload benchmarks of real large-scale cluster systems. We propose the notion of accumulated delay as a measurement of the performance of checkpointing techniques. It reflects the effect of checkpointing on system performance as experienced by the end users. Both theoretical analysis and empirical study show that tenant level checkpointing significantly outperforms traditional bulk checkpointing as measured by the amount of accumulated delays.