Programming-Level Power Measurement for GPU Clusters

GPU (Graphics Processing Unit) is becoming an integral part of mainstream computing systems with highly parallel, multithreaded, many-core processors of great computational power and high memory bandwidth. Finding the best tradeoff between performance and power efficiency is more challenging than mere performance tuning. To identify the principles of power-efficient programming on GPU clusters, we consider several primitives from the statements of parallel programs. The measured power for these primitives is applied to other programs for power estimation during program development. As a programming tool, this helps programmer evaluate the power issue of different optimization strategies.