Myths in power estimation with Performance Monitoring Counters

Many techniques have previously been proposed for using low-level CPU Performance Monitoring Counters in power estimation models. In this paper, we present some apparent myths regarding these techniques, and their potential impact. The underlying misconceptions include: (1) sampling rate and execution time can be left unspecified; (2) thermal effects are insubstantial; (3) memory events correlate well with power consumption; (4) compilation configuration does not need to be reported; and (5) metrics for performance evaluation of models are comparable. We aim to raise the awareness of these interesting issues, which existing power modeling techniques often do not address. Our discussions provide some guidance to avoid these myths and their effects through detailed specification of software and hardware configurations.