Knowledge-based checkpointing strategy for spot instances in cloud computing

Abstract

The Amazon EC2 offers spot-priced virtual machines (VMs) at a reduced price compared to on-demand and reserved VMs. However, Amazon EC2 can terminate these VMs anytime due to the spot price and demand fluctuation. Using spot VMs results in a longer execution time and disrupts service availability. Users can use fault-tolerant techniques such as checkpointing, migration, and job duplication to mitigate the unreliability of spot VMs. In this paper, a knowledge-based checkpointing strategy is proposed to minimize the overall checkpointing overhead during the execution of jobs. The proposed scheme uses real-time price history to decide when to take a checkpoint. Results show that the proposed approach can significantly reduce the turnaround time by 18% compared to Hourly Checkpointing Strategy and 9% compared to Rising-Edge Checkpointing Strategy. One can also achieve 54% to 78% reliability with a cost saving of 78% for the workload used with the described approach.