Power and Energy Normalized Speedup Models for Heterogeneous Many Core Computing

Continued technology scaling in VLSI has enabled more and more computation cores to be integrated in the same chip. This has facilitated the parallelization of processing and the increase of performance whilst keeping energy consumption at reasonable levels. To study the potential improvement of performance in such many core systems, three existing models have been popular in both the research community and industry. Amdahl's law is the original speedup model that estimates the maximum performance improvement with fixed workloads. Gustafson's law is a popular model that introduces variable workloads and estimates fixed time speedup. Sun and Ni combined the above two models into one considering the memory-bounded situation. These models are further extended via the Hill-Marty model to cover a limited form of heterogeneity. This paper extends these models to cover a more comprehensive assumption of core heterogeneity. We also present power and energy models based on the extended heterogeneous models. Our models cover popular power and performance control methods such as Dynamic Voltage Frequency Scaling (DVFS), power gating, etc. A case study is performed with an ARM big.LITTLE architecture containing Cortex A7 and A15 cores, including a comprehensive analysis with different ratios of parallel and sequential workloads to identify the most energy-efficient system configuration based on these models. Experimental results demonstrated high correlations between practically measured power normalized performance and that of the proposed extended models.