A Learning-Based Framework Incorporating Domain Knowledge for Performance Modeling
Energy efficiency has become the critical factor of computing performance in platforms from embedded devices, portable electronics to servers in datacenters. Power density and peak power demands increase in each generation of microprocessors, which directly lead to a higher operating temperature that exceeds the cooling capability available on current multi-core systems. These physical constraints seriously hinder the development of the next-generation computing platform. In this thesis, we propose Gray-Box computing, the methodology of a learning-based framework that incorporates the prior domain knowledge to quantitatively model every aspect of a multi-core system, including performance, power consumption and operating temperature. Experimental results show that the learned model achieves more than 96% accuracy, compared to actual industrial measurements or full-system simulations. By exploiting the learned model, the proposed Gray-Box computing has enabled a wide variety of applications from simulation speedup to multi-constrained optimization with respect to performance, energy efficiency and reliability for a multi-core system. Gray-Box computing has also been extended to model the performance specifically, the job inter-arrivals of a datacenter, which is in the scale of tens of thousands of cores. Experimental results are poised to demonstrate the strength of Gray-Box computing. Future work will focus on applying Gray-Box computing to model the usage dynamics of datacenters.