Sustainability - DSCoV Software Carbon Intensity

This repository contains some code to calculate the Software Carbon Intensity (SCI) score of a CPU job on our high performance computer OSCAR. This code is meant to be a starting point to developing new, more interesting manifests

Impact Framework

Impact Framework is a tool designed by the Green Software Foundation to calculate pipelines related to software sustainability. This workshop builds out a very rudimentary SCI score from an example job

Getting Started

Ensure you have a recent version of node. If you do not have node installed, please visit https://nodejs.org/en and follow the install instructions

Once node is installed, run npm install to install all the software packages

To run Impact Framework, use the npx command:

$ npx if-run --help

To run a pipeline in a given manifest file run:

$ npx if-run -m ${MANIFEST_FILE}

Manifests

Manifest files are artifacts that contain observations about a system and a pipeline to compute metrics about those observations. Each manifest in this repository is meant to illustrate different parts of the SCI score

01-basic-demo

This manifest file is there to show you the structure of a manifest file. This file takes observations and doubles the CPU utilization

02-operational-emissions

This manifest file calculates the operational carbon of the CPU from a job on OSCAR. The CPU for a given node is contained within the cpu-model.csv file in the resources directory. The regional grid intensity is from the EPA's most recent eGrid dataset (a copy is provided in the resources directory)

The pipeline takes the CPU utilization for the different observations and calculates the kilowatt-hour used based on the Thermal Design Power of a given CPU

03-embodied-emissions

This manifest file calculates the embodied carbon of the CPU by leveraging the SciEmbodied plugin from Impact Framework. The methodology used by the plugin is borrowed from the Cloud Carbon Footprint Dashboard

04-sci-score

This manifest file calculates the SCI score of the OSCAR job by combining the operational and embodied embodied emissions. The contents of the pipeline are copied from 02-operational-emissions.yml and 03-embodied-emissions.yml

Jobcarbon

Jobcarbon outputs yaml that is suitable for use as observations in an Impact Framework manifest. These observations are based on data in the Prometheus database. Prometheus is a timeseries database that stores information about recently run jobs

Jobcarbon requires an OSCAR job run within the last 30 days that took longer than a minute to complete

Installing

Copy the jobcarbon directory to OSCAR:

$ scp -r jobcarbon sshcampus.ccv.brown.edu:~/

Log into OSCAR and load the python module:

$ module load python/3.9

Install uv to manage the jobcarbon project:

$ pip3 install uv

Sync the jobcarbon project using uv:

$ cd jobcarbon && uv sync

Running

To run jobcarbon use uv:

$ uv run jobcarbon.py ${JOB_ID} ${START_TIME} ${END_TIME}

Copy the output of jobcarbon into a manifest file. Be careful about the indentation! Yaml is very picky. You'll still need to add the defaults and pipeline to the manifest alongside the inputs

The format of $START_TIME and ${END_TIME} is "YYYY-mm-ddTHH:MM:S". The exact parameters can be grabbed from sacct (NB: you only need the line where the job id has no suffix):

$ sacct -j ${JOB_ID} -o Job,Start,End