Mitigation Of Performance Variability Induced By Checkpoint-restart Induced By Dvfs

Abstract

As performance enhancement is accompanied by the aggressive integration of many-coresto a single chip and technology nodes approach deca-nanometer dimensions, the system’sfailure rate is becoming significant. Inevitably, computer systems must tolerate such failures.Both hardware and software methods are available enabling fault-tolerance to the systems.The Checkpoint/Restart technique provides reliability to the execution of an application.However, Checkpoint/Restart introduce an additional time overhead in order to achieve thefault-tolerance of the execution, that leads to performance variability.The scope of this thesis is to enhance a runtime manager, Depman, that orchestrates anapplication level Checkpoint/Restart technique so that such time overheads are absorbed,achieving performance predictability and reliability on the fly, by using Dynamic Voltage andFrequency Scaling (DVFS). A closed-loop implementation controlling the clock frequency isproposed, that quantifies the time overheads induced by the checkpoint restart process andadjusts the frequency levels of the CPU so that execution time converges to the normal.Depman was also modified to extend its portability to other platforms and applicationsand was tested using the self fault injection module to both the Intel’s Single-Chip CloudComputer (SCC) and an x86 general computing platform, evaluating both the executiontime and energy consumption of our scheme.