Performance and Power-Aware Classification for Frequency Scaling of GPGPU Applications

Abstract

The increased adoption of Graphics Processing Units (GPUs) to accelerate modern computational intensive applications, together with the strict power and energy constraints of many computing systems, has pushed for the development of efficient procedures to exploit dynamic voltage and frequency scaling (DVFS) techniques in GPUs. Although previous works have applied several pattern recognition techniques for GPGPU application classification, these approaches often result in many misclassifications when trying to identify which applications can benefit from DVFS. To circumvent this limitation, a new lightweight methodology for classifying GPU applications based on their performance and power consumption in the presence of GPU core frequency scaling is presented. The proposed methodology is based on a set of performance counters, such as memory bandwidth utilization and memory-related stalls, which are extracted during the application execution. Experimental results for a set of 20 applications from the Parboil, Rodinia and Polybench benchmark suites show that the proposed classification approach is able to correctly identify applications that can benefit from frequency scaling.