Transition Aware Scheduling: Increasing Continuous Idle Periods in Resource Units

Static power consumption has become a significant factor of the total power consumption in a system. Circuit level switching techniques reduce static power consumption by exploiting idle periods in processor components and placing them into low power modes or turning them off completely. In this paper, we propose a modified automaton-based list scheduling technique that augments the circuit level techniques by reducing the number of transitions between power modes, thereby increasing the of idle periods in resource units. Our scheduler uses a global resource usage vector and the usage vector of the last issued instruction(s) to select instruction(s) from the ready list such that resource units common to the last issued instruction and the selected instruction(s) are continuously active without creating a transition into low power mode. We estimate the power consumed in resource units using an energy model parameterized by the number of idle cycles, active cycles and transitions. We have implemented our algorithm in gcc and our simulations for different classes of benchmarks using an ARM simulator, with single and multi issue, indicate an average energy savings of 4%-15% in resource units with an average 0.25%-0.8% increase in the number of execution cycles