Abstract :
In this paper, we address the problem of efficiently managing the relative power demands of a high-performance GPU and its memory subsystem. We develop a management approach that dynamically tunes the hardware operating configurations to maintain balance between the power dissipated in compute versus memory access across GPGPU application phases. Our goal is to reduce power with minimal performance degradation. Accordingly, we construct predictors that assess the online sensitivity of applications to three hardware tunables-compute frequency, number of active compute units, and memory bandwidth. Using these sensitivity predictors, we propose a two-level coordinated power management scheme, Harmonia, which coordinates the hardware power states of the GPU and the memory system. Through hardware measurements on a commodity GPU, we evaluate Harmonia against a state-of-the-practice commodity GPU power management scheme, as well as an oracle scheme. Results show that Harmonia improves measured energy-delay squared (ED2) by up to 36% (12% on average) with negligible performance loss across representative GPGPU workloads, and on an average is within 3% of the oracle scheme.
