System Level Power Characterization of Multi-core Computers with Dynamic Frequency Scaling Support

Multi-core architecture has become prominent in modern processors including personal computers, large scale server systems and embedded systems. While multi core processors provide higher computing performance, they have higher power density and also consume more energy than traditional single core processors. Therefore, Dynamic Frequency Scaling (DFS) is mostly implemented in multi-core processors in order to reduce power and energy consumptions. However, the built-in dynamic frequency scaling ability in processors is not enough for highly dynamic and mixed workloads including computing intensive and I/O intensive computing workloads. In this paper we conducted extensive experiments to characterize the processor utilization, power consumption, and application performance. We provide the experiment results and observations. We found that: (1) There is no power model that is capable for different computing systems, even for the same workloads. (2) Processor utilization is not always the indicator for power consumption. For example, in some scenarios such as mixed type and I/O intensive applications, the power consumption is not proportional to the processor utilization and even keeps approximately unchanged. (3) Different computing systems with different subsystems configurations, i.e., processors, memories, and disks, have different power curves due to their proportion in total system-level power consumption because these different subsystems consume different basic power and dynamic power when running different workloads.