A mechanism to reduce energy waste in the post-execution of GPU applications

With the increasing demand of GPU accelerators for general purpose in HPC, the impact of energy consumption of these resources cannot be overlooked. To reduce the power consumption some strategies have been applied, but their approaches have been mostly focused on power savings during the application execution. This work focuses on post-execution energy savings. When the post-execution behavior is analyzed in newer GPU cards, it is observed that the power draw does not return to the idle state in an efficient way, creating an unexpected power waste. To overcome this inefficient return to idle and power draw waste in the post-execution, we developed a strategy to reduce the energy consumption considering a minimal impact on global performance. Using this strategy, we achieved energy savings up to 73 percent in the post-execution phase of a single run of a GPU application. In the case of sequential runs, the energy saving percentage depends of the waiting time gap between executions. (paper)