Adaptive Performance Sensitivity Model to Support GPU Power Management

Abstract

Integrated graphics units consume a large portion of power in client and mobile systems. Pro-active power management algorithms have been devised to meet expected user experience while reducing energy consumption. These techniques often rely on power and performance sensitivity models that are constructed at design phase using a number of workloads. Despite this, the lack of representative workloads and model identification overhead adversely impact accuracy and development time, respectively. Conversely, two main challenges limit runtime post-design identification: the absence of sensitivity feedback from the system and the limited computational resources. We propose a two-stage methodology that first identifies the features of the sensitivity model offline by leveraging a reduced amount of training data and then uses recursive least square algorithm to fit and adapt the coefficients of the model to workload changes at runtime. The proposed adaptive approach can reduce offline training data by 50% with respect to full offline model identification while maintaining accuracy as much as 95% on average.