System-level power-performance efficiency modeling for emergent GPU architectures

The authors propose separate system-level power and performance models of modern GPU based systems that can be integrated to better understand the relationships between power and performance at scale. Our runtime hardware counter driven approach combines an analytical performance model for GPU based systems with a machine-learning-based artificial neural network (ANN) power model to form a power-performance efficiency model. Specifically, with the proposed models, users can: (1) Predict optimal configurations to maximize efficiency for single GPU and GPU-based clusters; (2) Explore and identify application and system bottle- necks for performance, power, and energy consumption, and explain the root causes. The experimental results show that our models are accurate for predicting power (within 2.1%) and performance (within 6.7%) for representative CUDA kernels running on Fermi based GPUs. Furthermore, results indicate that capturing non-linear effects between power and performance leads to improved accuracy over the best available GPU power models-MLR based approaches. Our models can also be simply used by programmers to identify power-performance bottlenecks. We apply our combined model to suggest the optimal number of GPU based cluster nodes for SHOC benchmarks from ORNL to achieve the best energy-performance efficiency on the Keeneland supercomputer.