The GSF’s SCI standard specifies that energy consumption should encompass all energy consumed by reserved or provisioned hardware, not solely the hardware used while an application is running. This includes idle power, encompassing not only static power but also the energy consumed by items necessary to be functional for the application to run such as the kubelet or the control plane.

According to theSPECpower benchmark, active idle power consumption can vary widely, ranging from 20% to 60% of the power consumed at maximum utilization. In absolute terms, this ranges from 100Watts to 1300Watts, equivalent to the power consumption of a classic light bulb or an air conditioner in a household (although there are differences in carbon emissions). The idle power can be even higher when considering GPU power consumption.

The differentiation between idle and dynamic power has been widely investigated in the research literature[1].

It is crucial to isolate idle power from dynamic power for reasons including fair attribution of power consumption to applications and to help create more accurate power models.

Dynamic energy consumption is directly correlated to resource use. On the other hand, idle power is the base power consumption that occurs regardless of resource utilization. A power model that uses resource utilization to estimate power consumption will only accurately reflect dynamic power. Idle power, is a constant power consumption that must be attributed to all running applications.

.

Suppose we have a server with a linear power model for a 16-core server X, described by a simple formula as simple as below where Idle=200 and Coeff=8:

Power Consumption (P) = Idle Power (Idle) + Dynamic Power Consuming Factor (Coeff) * Resource Utilization (x)

In the third scenario, isolating idle power enhances predictability and clarity. This method divides constant power (unaffected by resource use) and defines dynamic power solely based on resource utilization. Specifically, according to the power model, Client A’s dynamic power is 80W. By allocating idle power based on resource ratio and absolute power, Client A receives 40W of idle power, while Client B gets 160W. and absolute power, Client A receives only 40W of idle power, leaving Client B with 160W of idle power.

These unequal distributions of idle power highlight unfairness because idle power consumption isn’t tied to resource utilization. Applying the power consumption ratio based on resource requests doesn’t accurately reflect actual dynamic power consumption corresponding to real resource utilization.

Estimating idle power in virtual private clouds is challenging due to unknown physical specs and the number of tenants (VMs). To address this, we’ve developed a training framework focused on extracting and training the dynamic power model alone. This approach reduces dependency on unknown factors and enhances energy-efficient optimization for virtual machines[4].

Complete waste accounting at the user level remains a challenge, necessitating further research and collaboration within the industry. Transparent, fine-grained energy metrics for public cloud efficiency remains important, and needs to have challenges addressed such as idle power attribution and estimation. Tools like Kepler pave the way for sustainable cloud computing practices.