Efficient calibration of thermal models based on application behavior

Abstract

With increasing power densities, raising operating temperatures in chips threaten system reliability. Thermal control therefore has emerged as an important issue in system design and management. For dynamic thermal control to be effective, predictive thermal models of the system are needed. Such models typically use power as input, which renders them difficult to use in practical systems, where power monitoring is not available at processor or chip level. In this paper, we describe a methodology to infer the thermal model based on the monitoring of existing temperature sensors and of instruction counter registers. This allows the thermal model to be easily established, calibrated, and recalibrated at runtime to account for different thermal behavior due to either variations in fabrication or to varying environmental parameters. We validate the proposed methodology through a series of experiments. We also propose and validate an extension of the model and associated methodology for multicore processors.