sustainability-measurement-agent 0.2.1

Sustainability Measurement Agent - A tool for collecting and analyzing sustainability metrics in cloud native enviroments.

Sustainability Measurements Agent (SMA)

|

SMA is an open-source tool designed to help deploy, collect and report sustainability measurements on cloud-native applications. It is not itself a measurement tool, but rather a framework to orchestrate, combine and aggregate results from emerging suite of sustainability measurement tools.

Capabilities (planned)

• Configuration driven operation

• Support for multiple scenarios
  • Continuous monitoring and reporting (as a Service)
  • Experiment Measurements / Benchmarking
  • Ad-hoc measurements
  • Programmatic measurements via API
• Multiple report formats
  • CSV
• Post processing and aggregation
  • Pandas
  • Grafana Dashboards
• CLI tool

Use / Install

Make sure you have Python 3.8+ as well as hdf5 and jupyter installed. You may need to install additional dependencies depending on your setup, e.g., c-blosc lzo bzip2 gcc.

pip install sustainability-measurement-agent

touch main.py

import sys
from sma import SustainabilityMeasurementAgent, Config, SMAObserver

config = Config.from_file("examples/minimal.yaml")
sma = SustainabilityMeasurementAgent(config)
sma.connect()
def wait_for_user() -> None:
    input("Hit Enter to to Stop Measuring...")

sma.run(wait_for_user)
sma.teardown()
report_location = config.report.get("location")
if report_location:
    print(f"Report written to {report_location}")

$ python main.py
$ Hit Enter to to Stop Measuring...

SMA also ships with a cli tool, that allows some basic operations:

$ sma --help
Usage: sma [OPTIONS] COMMAND [ARGS]...

Options:
  --help  Show this message and exit.

Commands:
  fetch  Attempts to fetches measurements again using the provided
         configuration file.
  list   Lists all reports for a given configuration file.
  run    Run the Sustainability Measurement Agent with a given...

Architecture and Lifecycle

SMA follows a simple lifecycle:

Fully driven by configuration files, SMA connects to external services (e.g., Prometheus), deploys measurement tools (e.g., Kepler), and once instrcuted via the API (run()), starts the observation according to the configured mode (trigger, timer, continuous). Finally, SMA collects the measurement data, generates reports, and tears down any deployed infrastructure (if configured to do so).

SMA exposes an observer API (SMAObserver), which allows to hook into the lifecycle events (see the diagram) to extend SMA's functionality programmatically, or to integrate into experiment frameworks.

Modules can optionally be used to add functionality, such as installing monitoring infrascture, preparing deployments, running workloads, and enhancing reports.

Configuration

SMA is configured via YAML files. See the examples/ folder for sample configurations. The configuration allows to specify which measurement tools to use, how to deploy them, and how to collect and report the measurements.

Configuration Structure

The configuration file consists of the following main sections:

1. Version

sma:
  version: 0.2.0

Specifies the SMA configuration schema version.

3. Observability Services

services:
  prometheus:
    address: http:/<IP>

Defines external services used for data collection, sofar we only support Prometheus.

Field	Description
prometheus.address	URL of the Prometheus server for metrics collection

5. Observation

observation:
  mode: trigger 
  window:
    left: 10s   
    right: 10s 
  max_duration: 10s
  targets:
    - name: sut
      namespace: kubeai

Field	Options/Type	Description
mode	trigger, timer, module	trigger: Measurement starts/stops based on a Python function that blocks timer: Measurement based on time window settings\
module: run a trigger function from a module, e.g. a workload runner
trigger_module	When in module mode, specify the module name to call the trigger from
window.left	Duration (e.g., 10s)	Time to capture before the measurement trigger
window.right	Duration (e.g., 10s)	Time to capture after the measurement trigger
max_duration	Duration (e.g., 10s)	Optional total measurement duration, before a run is interuppted.
targets[].name	String	Name identifier for the target workload
targets[].namespace	String	Kubernetes namespace of the target workload

Both left and right windows are optional and mean, SMA will wait before handing off the measurement to the next API call. The max_duration is optional and specifies the maximum time to run the measurement, before it is interrupted. For the time trigger mode, the treatment will be stopped after the max_duration has elapsed (so expect to see a warning in the logs if the treatment is interrupted).

6. Measurements

measurements:
  # Custom PromQL query
  - measurement_name:
      type: aggregate
      layer: pod         
      query: <PromQL query>
      step: 30           
      unit: watts       
      target:
        - all        

Field	Options/Type	Description
type	raw, aggregate, counter, guage	raw: Unprocessed metric data aggregate: Aggregated/computed metrics
layer	substrate,pod, node, process	The abstraction level being measured
query	PromQL string	Prometheus query to retrieve the metric. Supports template variables: ${namespace}: Replaced with target namespace ${SMA_SELECTORS}: Additional label filters
sensors	List of strings	List of sensors required for this measurement (used with prepared_query)
step	Integer (seconds)	Sampling interval in seconds
unit	String	Measurement unit (e.g., watts, joules, cpu_time_seconds, metadata)
target	List of strings	Target names to measure, or all for all defined targets

Query Methods:

• Custom Query: Use query field with custom PromQL
• Prepared Query: Use prepared_query field with predefined SMA queries (e.g., pod_kepler_energy_consumption, node_scaphandre_energy_consumption)

Example Measurements:

• CPU Usage: Container CPU usage across namespaces
• Wall Power: Physical power consumption from smart plugs (e.g., Shelly)
• Pod Power (Kepler): Per-pod energy consumption via Kepler
• Node Power (Scaphandre): Host-level power measurements
• Container Info: Metadata about running containers

7. Report

report:
  format: csv          # Output format
  location: reports/${startTime}_${runHash}/
  filename: "${name}_${startTime}.csv"

Field	Type	Description
format	String	Output format (currently supports csv)
location	String	Directory path for reports. Supports template variables: ${startTime}: Timestamp when measurement started ${runHash}: Unique hash for the measurement run
filename	String	File naming pattern. Supports template variables: ${name}: Measurement name ${startTime}: Measurement start time

Modules can add their own files to the report directory through the onReport event.

API

Development

1. Install pipenv: pip install pipenv
2. Install dependencies: pipenv install --dev
3. Activate virtual environment: pipenv shell
4. Do your thing. Eventually with tests.
5. If you add new dependencies, run pipenv lock --pre to update the Pipfile.lock.
6. If you want to release a new version, update the version in pyproject.toml and run pipenv run pyproject-pipenv --fix and commit the changes.
7. Tag your release: git tag vx.y.z and git push --tags
8. Create a new release on GitHub, this will trigger the upload to PyPI via GitHub Actions.

Acknowledgements

This project builds upon prior work done in the OXN project, GOXN project and CLUE and is part of the research at ISE-TU Berlin.