Power Modeling for Effective Datacenter Planning and Compute Management.

摘要: Datacenter power demand has been continuously growing and is the key driver of its cost. An accurate mapping compute resources (CPU, RAM, etc.) hardware types (servers, accelerators, to consumption emerged as a critical requirement for major Web cloud service providers. With global growth in datacenter capacity associated consumption, such models are essential important decisions around design operation. In this paper, we discuss two classes statistical designed validated be accurate, simple, interpretable applicable all configurations workloads across hyperscale datacenters Google fleet. To best our knowledge, largest scale modeling study kind, both scope diverse planning real-time management use cases, well variety workload used validation. We demonstrate that proposed techniques, while simple scalable, predict with less than 5% Mean Absolute Percent Error (MAPE) more 95% Power Distribution Units (more 2000) using only 4 features. This performance matches reported accuracy previous started-of-the-art methods, significantly features covering wider range cases.