Investigating the Effects of Task Scheduling on Thermal Behavior

 Thermal behavior is expected to be an important design consideration for the next generation of microprocessors. Most dynamic thermal management techniques are based on the principle of ‘reactive throttling’, where some form of performance throttling (such as voltage or frequency scaling) is invoked, after a pre-architected temperature threshold has been exceeded. As such, there is performance degradation with each such DTM technique. In some cases, if the temperature threshold is set to a conservative value for package/cooling cost considerations, the degradation can be quite severe. In this paper, we investigate the potential benefits of temperature-aware task scheduling with minimum or no performance overhead. We explored policies based on metrics such as average temperature; dynamic profiling of threads and localized heating, instead of IPC-based metrics. Our preliminary results indicate that thermal-aware task scheduling provides significant alleviation on chip temperatures. We observed up to 52% reduction in the number of cycles above the thermal threshold, compared to a thermally-oblivious policy. On average, our MinTemp policy yields less than 3% thermally critical cycles, even on a challenging benchmark set such as SPEC2000 with 12 of the 25 benchmarks above 360K. As we employ already existing scheduling infrastructure, there was no considerable change in net performance.