Energy-efficient checkpointing in high-throughput cycle-stealing distributed systems

Abstract

Checkpointing is a fault-tolerance mechanism commonly used in High Throughput Computing (HTC) environments to allow the execution of long-running computational tasks on compute resources subject to hardware or software failures as well as interruptions from resource owners and more important tasks. Until recently many researchers have focused on the performance gains achieved through checkpointing, but now with growing scrutiny of the energy consumption of IT infrastructures it is increasingly important to understand the energy impact of checkpointing within an HTC environment. In this paper we demonstrate through trace-driven simulation of real-world datasets that existing checkpointing strategies are inadequate at maintaining an acceptable level of energy consumption whilst maintaing the performance gains expected with checkpointing. Furthermore, we identify factors important in deciding whether to exploit checkpointing within an HTC environment, and propose novel strategies to curtail the energy consumption of checkpointing approaches whist maintaining the performance benefits.