2024.05.21

[seanmcilroy29]

---

2024.05.21 Agenda/Minutes

---

Time 0800 (PT) / 1600 (BST) - See the time in your timezone

• Chair – Adrian Cockcroft
• Chair - Pindy Bhullar (UBS)
• Convener – Sean Mcilroy (Linux Foundation)

Antitrust Policy

Joint Development Foundation meetings may involve participation by industry competitors, and the Joint Development Foundation intends to conduct all of its activities in accordance with applicable antitrust and competition laws. It is, therefore, extremely important that attendees adhere to meeting agendas and be aware of and not participate in any activities that are prohibited under applicable US state, federal or foreign antitrust and competition laws.

If you have questions about these matters, please contact your company counsel or counsel to the Joint Development Foundation, DLA Piper.

Recordings

WG agreed to record all Meetings. This meeting recording will be available until the next scheduled meeting.

Roll Call

Please add 'Attended' to this issue during the meeting to denote attendance.

Any untracked attendees will be added by the GSF team below:

• Full Name, Affiliation, (optional) GitHub username

Agenda

• Approve agenda
• Approve previous Meeting Minutes

Renewable Energy Percentage

• Completed defining language for the column headers
  Action Item: Members requested to assist in defining column headers GSF Real-Time Cloud Renewable Energy Percentage spreadsheet
• RTC Cloud Region Impact Framework Documentation

Discussion Board

• Review miro dataflow in detail
  • Continue to review and approve the proposed Scope.

Open Issues Project Board

Open Issues for review

• Merge the various versions of cloud region metadata #40
• How should GPU energy use be estimated? #37
• Fill out geolocation column for Cloud Region impact framework #31
• Fill out Watttime column for Cloud Region impact framework #32
• Fill out Electricity Maps data for Cloud Region impact framework dataset #33
• Decide on source and fill in energy production data for Cloud Region impact framework dataset #34
• Complete CFE region and any other data gaps in Cloud Region dataset #35
• AWS PUE Data Alignment #29
• GCP PUE Data Alignment #28
• Azure PUE data alignment #27
• Cloud provider Power Purchase Agreement (PPA) generation locations as an Impact Framework dataset #20
• Windows Energy Estimation Tool #19
• What are the use cases for real time power and carbon intensity information? #12
• How can energy data collection and reporting be implemented by a cloud provider? #4

AOB

• Topics added during the meeting

Next Meeting

• 04 June

---

Action Items

• Contribute additional relevant projects, tools, and data sources to the mapping diagram.
• Add links from the discussion to the mapping diagram on the real-time cloud GitHub mirror.
• Follow up with Microsoft to clarify the mapping of data centres to cloud regions in the impact framework dataset.
• Members - Review issues on GitHub and provide comments/input.

[mrchrisadams]

Attended

[adrianco]

Attended

[coopere]

attended

[jawache]

Attended

[PindyBhullar]

Attended

[seanmcilroy29]

MoM

Adrian opens the meeting at 0800 (PT) / 1600 (BST)

Integrating data from a spreadsheet into a metadata system for cloud providers.
Adrian highlights the need for Microsoft data and metadata and wants to merge geolocation data. Adrian and Asim also discussed integrating data from a spreadsheet into an impact framework pipeline. The team is working on developing a plugin to surface spreadsheet data into an impact framework, which will help avoid constant syncing issues.

Cloud metadata and GPU support.
Asim and Adrian discussed the cloud metadata impact framework model, which currently has a modified CSV snapshot.
They consider whether to remove the spreadsheet from the model or keep it as part of the cloud metadata. Adrian and Asim discussed using the CCF tool for data collection, but it has maintenance issues.

TDP curves and GPUs for cloud providers.
Adrian is in discussions with cloud providers to obtain TDP curves for their instances, as it is difficult to obtain actual wattage data. They have observed that Intel chips sold to cloud providers have different TDP curves than those sold to consumers, even for large consumers. Asim and Adrian are talking about the efforts made by cloud providers to improve energy efficiency and transparency. They are also discussing GPU power consumption and data sources in their video meetings.

Measuring energy consumption of GPUs in cloud computing.
Asim and Adrian discussed the challenges of measuring the energy consumption of GPUs in the cloud. Adrian mentions that cloud providers block GPU metrics to prevent security leaks. Users typically measure power consumption to ensure they get maximum FLOPs, as it's easier to measure power than throughput in large-scale GPU use cases. Adrian and Asim also discuss the impact framework and hackathon. Asim shares his learnings from the hackathon related to measuring the carbon impact of AI workloads. Asim highlights two case studies and a plugin registry as potential solutions to the problem of understanding the energy consumption and carbon impact of AI workloads.

Improving carbon emissions data for software development.
Asim and Adrian discussed challenges with error handling in a pipeline, including issues such as "undefined is not a function" and empty spots. Teams are currently working on diagnosing issues and identifying reasons for pipeline failures, such as missing data or complex moving parts. Adrian plans to add a column to the data set for water usage efficiency, as per a request a member request. Asim suggests using litres of water per kilowatt hour as the unit for water usage efficiency and agrees with Adrian's decision to add the column. Adrian also discusses the progress and recent updates of the Kepler project.

Carbon footprint monitoring and energy efficiency in cloud computing.
OpenNebula is developing models for fine-tuning machine learning models based on CPU and GPU types to improve benchmarking confidence. The group discusses using carbon footprint data for cloud workloads in real-time and using reference implementations for energy monitoring and carbon footprint estimation. Asim suggests publishing a spreadsheet or CSV to make the data more accessible and incentivize others to fill in gaps.

Carbon emissions data for cloud computing workloads.
Adrian and Asim discussed a modified spreadsheet version that includes additional Azure information. They are trying to merge this modified version back into the original. Their goal is to use the data for a workflow in a CI pipeline, which includes carbon intensity numbers, energy maps, and power usage efficiency numbers. Adrian explains how cloud providers' carbon footprint is calculated, including using carbon-free energy percentages and location models. Adrian also notes that country-level energy numbers are in a spreadsheet, but the market model does not account for marginal carbon emissions.

Using AI for power consumption analysis in data centres.
Members discussed using models for power consumption analysis, focusing on CPU, DRAM, and GPU usage.
Group discussed Kubernetes, containers, and process levels, with some confusion around documentation and installation processes. Adrian mentions that some organizations are actively running workloads on GPUs, such as training and HPC batch jobs, and wonders if there's someone they could talk to for data collection or insights. Adrian also mentions that the AI architecture is shifting towards GPU centricity, with larger memory domains and shared memory buses and that Blackwell systems will come out later with even bigger capabilities. Adrian needs help cleaning up a forked data set and scheduling an asynchronous meeting with Microsoft.

Action Items

• Sync up Adrian with Microsoft contact to discuss the forked dataset and gaps in their data
• Publish the initial dataset on GitHub for community input
• Investigate opportunities for Kepler models to be exposed and utilized more broadly
• Schedule a follow-up meeting between Adrian and Microsoft contact