Using performance counters for runtime temperature sensing in high-performance processors

As energy consumption in high-performance systems has increased, thermal management has become a big challenge. Providing a cost-effective and detailed temperature sensing mechanism is crucial to effectively employ a thermal management technique. Existing hardware sensors are too costly to implement and add additional heat while software simulations fail to account for all possible hardware effects. In this paper, we describe a software solution for temperature sensing that uses real hardware resources such as performance counters. The resulting temperature model provides a detailed spatial gradient of the processor and executes at runtime. In particular, the model is configured for the Pentium 4 processor. We run SPEC2000 benchmarks to analyze the thermal behavior of applications and explain the potential benefits of using our model for temperature-aware research.