Getting Energy Consumption Per Process

[Saicasm]

Hi everyone,

I was wondering if it is possible to get energy consumption metrics per process. Currently, the output from the CSV file provides general metrics, but I’d like to retrieve energy consumption for a specific process, such as nginx.

Is there a way to obtain per-process energy consumption using RAPL?
If not, would it be possible to achieve this within a Kubernetes cluster?
Any insights or alternative approaches would be greatly appreciated. Thanks!

[TheElectronWill]

Hello!

RAPL only provides hardware-level information. A processing (transform) step is necessary to compute an estimate of the energy consumed by software components (processes, pods, jobs, whatever) by combining their CPU usage and RAPL energy consumption.

This is currently a work in progress. We have a first version of the basic bricks, but there are some caveats.
Both process attribution and K8S pod attribution are supported.

What you need to do:

1. enable 3 plugins

   • rapl
   • energy-attribution
   • a plugin that collects cpu usage of processes (procfs) or pods (k8s)

2. configure energy-attribution to use the metrics that correspond to your use case.

Example for K8S pods:

[plugins.energy-attribution]
# the metric that contains the energy consumption of the hardware
consumed_energy_rapl = "rapl_consumed_energy"
# the metric that contains the hardware usage
hardware_usage_cgroup = "cgroup_cpu_usage_user"

We will provide docs, containers and helm charts in the near future :)

[Saicasm]

Hey @TheElectronWill , thank you for responding. Is there a way to enable energy attribution? I tried the default way of activating it along with the other plugins, but it didn't work. Also, if I am not wrong, this is the flow for getting the process-level
metrics, right? Please correct me if I am wrong. I would love to contribute to either completing the documentation or even help with the development. If you could point me in the right direction, I can take off from there. Thank you!

FYI: This is how I started the agent : ./target/debug/alumet-agent --plugins rapl,csv,procfs,energy-attribution

[TheElectronWill]

For now you need to manually add the energy-attribution plugin to the agent by

• adding it as a dependency of the agent
• modifying main.rs
• recompiling the agent

It is not included by default, since it's not ready yet ^^ We already have someone working on that part, but it's great that you want to help! We can find something you could do by discussing your use case. You can also look at the existing non-assigned issues.

I agree with your diagram. Nitpick: energy-attribution is a plugin too, and you can enable multiple transforms :)

[Saicasm]

Hello @TheElectronWill,

Thank you for clarifying and directing me towards the right steps. I was able to achieve pod-level metrics with energy attribution to each pod. Here are the steps for anyone who might want to try it out:

Steps to Set Up Pod-Level Energy Attribution

1. Update and Build the Alumet Agent

Modify Cargo.toml

1. Navigate to the agent folder
2. Add the energy attribution plugin under Linux-only dependencies:

plugin-energy-attribution = { path = "../plugin-energy-attribution" }

Modify main.rs

Add the plugin to the Linux-only plugins list:

#[cfg(target_os = "linux")]
{
    plugins.extend(static_plugins![
        plugin_socket_control::SocketControlPlugin,
        plugin_cgroupv2::K8sPlugin,
        plugin_cgroupv2::OARPlugin,
        plugin_oar2::Oar2Plugin,
        plugin_rapl::RaplPlugin,
        plugin_perf::PerfPlugin,
        plugin_procfs::ProcfsPlugin,
        plugin_nvidia::NvidiaPlugin,
        plugin_energy_attribution::EnergyAttributionPlugin,
    ]);
}

Build the Agent

cargo build -p alumet-agent

2. Update alumet-config.toml

[plugins.csv]
output_path = "alumet-output.csv"
force_flush = true
append_unit_to_metric_name = true
use_unit_display_name = true
csv_delimiter = ";"

[plugins.rapl]
poll_interval = "1s"
flush_interval = "5s"
no_perf_events = false

[plugins.energy-attribution]
consumed_energy_rapl = "rapl_consumed_energy"
hardware_usage_cgroup = "cgroup_cpu_usage_user"

[plugins.k8s]
path = "/sys/fs/cgroup/kubepods.slice/"
poll_interval = "1s"
kubernetes_api_url = "https://128.164.171.96:6443/"
hostname = ""
token_retrieval = "kubectl"

3. Verify Kubernetes & Run the Agent

ls -l /sys/fs/cgroup/kubepods.slice/
./target/debug/alumet-agent --plugins rapl,k8s,energy-attribution,csv

In addition, I wanted to mention a potential issue I've noticed. In the CSV output I'm receiving, I see that the pod information is displayed as Pod UUID rather than the pod name that one would get from kubectl get pods. I'm not sure if this is intended behavior, a work in progress, or possibly a bug, as I noticed there are some functions in the utils.rs file that seem to handle similar functionality. I understand this feature may still be under development, but I thought it might be helpful to highlight this observation in case it's unintended behavior

metric	timestamp	value	resource_kind	resource_id	consumer_kind	consumer_id	name	namespace	node	uid	__late_attributes
cgroup_memory_anonymous_B	2025-03-28T02:48:17.463295324Z	1122304	local_machine		cgroup	/sys/fs/cgroup/kubepods.slice/kubepods-besteffort.slice/kubepods-besteffort-pod8435bfbf_d50a_4644_a126_89e51f9daf3d.slice/memory.stat				pod8435bfbf_d50a_4644_a126_89e51f9daf3d
cgroup_memory_file_B	2025-03-28T02:48:17.463295324Z	0	local_machine		cgroup	/sys/fs/cgroup/kubepods.slice/kubepods-besteffort.slice/kubepods-besteffort-pod8435bfbf_d50a_4644_a126_89e51f9daf3d.slice/memory.stat				pod8435bfbf_d50a_4644_a126_89e51f9daf3d

The uid field contains values like pod8435bfbf_d50a_4644_a126_89e51f9daf3d instead of human-readable pod names.