DRAM metrics not supported in 0.10.0 post rewrite? kepler_container_joules_total missing

[MA3CIN]

Target Version

Current version (0.10.0+) - New Architecture

Feature Description

Comparing documentations of versions 0.9.0 and 0.10.0, as well as their deployments on my clusters, the archive version has more metrics per container and process than the newest one.

I'm specifically talking about the "kepler_container_joules_total " metrics, which included "CPU, dram, gpus, and other host components" ( 0.9.0 docs - https://sustainable-computing.io/archive/design/metrics/#kepler-metrics-for-container-energy-consumption).
The newest version only supports cpu related metrics for containers - kepler_container_cpu_joules_total (0.10.0 - https://sustainable-computing.io/kepler/design/metrics/)

I've read the Slack announcement, including the new "limitations" section, which talks about only supporting the RAPL/powercap framework. I've also read the Slack comments regarding the "past models being way off on power consumption"

Will the DRAM related metrics be back eventually in newer Kepler versions?

Problem Statement

Less metrics available than before in 0.10.0 (specifically dropping DRAM support)

Proposed Solution

A comment regarding DRAM/GPU power metrics's place in the Kepler Roadmap

Alternatives Considered

No response

Additional Context

No response

[vprashar2929]

@MA3CIN Thanks for raising this. Quick rundown of what changed and why:

In the archived version (e.g., v0.9), Kepler exposed separate, static metrics for different components (like package, core, dram), sometimes even when the underlying hardware support was missing. This could lead to misleading data, as you noted, where metrics were produced but were either empty or based on estimations that might not reflect reality.

In the current version (v0.10.0+), we have shifted to a more dynamic approach that strictly reflects the available hardware sensors (RAPL). Instead of separate metric names for each component, we now use a single metric kepler_container_cpu_joules_total with a zone label.

This zone label is populated dynamically based on the RAPL domains available on your specific machine:

• If your machine supports Package and Core power domains, you will see zone="package" and zone="core".
• If your machine also supports DRAM power domain, you will additionally see zone="dram".

This design ensures that Kepler only reports energy consumption for components that are actually being measured by the hardware, providing a more accurate and "honest" picture of your system's energy usage. If you are not seeing dram in the zones, it means the RAPL interface on your node is not exposing a DRAM power domain.

[MA3CIN]

Hi @vprashar2929 , thanks for weighing in on this. Hopefully this goes for both container and process versions of Kepler, I will tinker with this shortly :)

[vprashar2929]

@MA3CIN Are we okay to close this, or is there more you’d like to follow up on?