How is PowerAPI intended to be used for energy attribution within VMs today?

[Louis1870]

We are currently working on energy attribution inside virtual machines using PowerAPI and I would like to ask which approach is currently recommended for this use case.

There seem to be two separate aspects which are currently unclear to us.

On the model/formula side:
it is unclear to us whether SmartWatts or VirtualWatts should currently be used for energy attribution inside virtual machines (see https://powerapi.org/getting_started/ and https://github.com/powerapi-ng/virtualwatts-formula).

On the sensor/input side:
from our understanding, the procfs sensor used by older VirtualWatts setups appears outdated/deprecated, while the current recommendation seems to be the HWPC sensor instead.

According to the documentation, the HWPC sensor appears to run on the host system and not inside the VM itself.

This raises the question:
which model should currently be used inside the VM, with which sensor/input source, and how is the model expected to obtain/read the required monitoring data inside the VM?
Any clarification or updated references/examples would be greatly appreciated.

[rouvoy]

Thanks for reaching out @Louis1870 🙂

VirtualWatts (also known as BitWatts) is the proof of concept of a formula we published in Eurosys'12 and implemented atop PowerAPI.

No matter the implementation, the core idea was to use a formula on the host server that estimated in real-time the power consumption of each host VM and forwarded this information inside the VM by using a Virtio interface.

This information was then read by another instance within each VM, which could combine it with ProcFS to further distribute the power consumption of the VM among the processes running inside this VM.

The current implementation is mostly outdated because PowerAPI keeps evolving since we published this work, while the requirements for such a virtual-formula was not that strong in the end. Nonetheless, I think you can easily port it to the latest version of our tools if you see some opportunities to extend this work.

All the best,
Romain