Software Performance Engineering Considerations in Unreliable Computing Environments

For some computing systems failure is rare enough that it can be ignored. In others, failure is common and software design considerations can have a significant impact upon the performance of applications running on the systems. In this paper, we discuss software design issues that should be considered whenever jobs execute in unreliable computing environments. Specifically, we show that if proper checkpointing mechanisms are not properly implemented, then under certain conditions completion times of applications executing on the system exhibit properties of heavy-tail or power-tail distributions, which can lead to unpredictable and long completion times.