Invasive Computing for Power Corridor Management

Abstract

This paper investigates the use of invasive computing to enforce the power budget in an HPC infrastructure. Invasive MPI along with the Invasive Resource Manager (IRM) provides an infrastructure for developing malleable/invasive applications. In IRM, a power model is used to predict the power consumption of each application. If a violation in power corridor is predicted, IRM reconfigures the node allocation among the applications to keep the whole system back into the power corridor. Since development of invasive applications is a complex task, a new programming model called Elastic Phase Oriented Programming (EPOP) is developed to simplify the invasive programming. This model is also capable of collecting and sharing power usage metrics as well as performance metrics to IRM.