Multi level Kepler Deployment

[tawalaya]

Hi there,
I'm looking into deploying Kepler in a KVM cluster that will host multiple Kubernetes nodes. Could you give me any guidance on whether this is possible and, if so, how to do it?

In the documentation, the possibility to do so is hinted at. (VM with node info and power passthrough from BM ) I'm guessing that the RPM package needs to be installed on the KVM hosts, and the VMs need to mount specific files. But I couldn't find any precise doc on that.

Thanks in advance :)

[rootfs]

Kepler can run inside of VMs and use ML models to estimate power consumption. The VMs don't have to have VM passthrough (actually the RAPL counters won't be on VMs anyway).

[tawalaya]

Yes, I understand, but my question was. What if I wanted to use passthrough? I've seen models like scaphandre where you can pass a custom rapl counter through to the VM that only accounts for the usage of the VM. Is something like this possible? Alternatively, what files is Kepler looking for to read RAPL counters? Could a combination scaphandre BM + VM Kepler + Kubernetes work?

[aharivel]

@tiwatsuka Actually it has landed in Qemu just yesterday:
https://gitlab.com/qemu-project/qemu/-/commit/0418f90809aea5b375c859e744c8e8610e9be446

[tiwatsuka]

Thank you @aharivel! It's great that Kepler on VM can obtain components power from RAPL.

Having said that, I believe multi level Kepler deployment still useful to collect other estimated VM power metric such as platform power from ACPI or Redfish. I know that I can use power estimation models, but I prefer measurement approach in terms of its explainability and accuracy without preparation.

[aharivel]

@tiwatsuka I understand your concern and this kind of decision is out of my hand.

However, in my opinion, getting power measurement from IPMI or ACPI is not really handy for what Kepler tries to achieve.
Kepler is not just here to export raw data, it processes them before exporting them. RAPL has the smallest granularity in terms of power plane (Package, Core, Dram,...) and it's the most convenient for calculating the ratio of each processes.
Using platform power would include power usage of disks, fans, others hardware etc. and this is not really possible to know exactly which processes has triggered the fan, why the disk are spinning more at this points, and so on. I guess you get my point.

In the end, if you need to have the global power consumption of your system to add more visibility of your power usage, adding a IPMI / Redfish prometheus exporter is just as easy as running Kepler. However if you need to target which processes is the most power hungry, Kepler is way more convenient for that matter.

[tiwatsuka]

@aharivel I understand your point clearly. My concerns may stem from my specific use case. I am attempting to use Kepler not only to understand the trend of energy consumption but also to calculate the greenhouse gas emissions of each pod, in accordance with standards such as the GHG Protocol. I’m uncertain whether the electricity consumed by components like fans or disks can be disregarded even if it’s unclear which process is responsible for this consumption. Therefore, it would be beneficial for me if the estimated total energy consumption of each pod on VM could be presented in an explainable manner.

[aharivel]

@tiwatsuka Did you already have a look at this blog:
https://deploy-preview-435--tag-env-sustainability.netlify.app/blog/2024-06-idle-power-matters-kepler-metrics-for-public-cloud-energy-efficiency/
Maybe you could find answer in it ?