Green computing metrics infrastructure for GreenDIGIT - Assessment of EGI tools

Tuesday Feb 4, 2025, 10:00 AM → 11:30 AM Europe/Amsterdam

Starting_point_for_discussion

Meeting Summary for Green computing metrics collection infrastructure for GreenDIGIT

Feb 04, 2025 10:00 AM Budapest ID: 638 4116 8170

 

Minues initially generated by Zoom AI companion, then reviewed by Gergely Sipos (EGI F.)

Quick recap

• The meeting was among EGI members to discuss tools for the possible use as metrics infrastructure in GreenDIGIT. The team discussed the characteristics of metrics that have to be stored/shared to compare environmental efficiency of sites, and explored the EGI Configuration Mgm System (GOCDB), the EGI Accounting System (From STFC and CESGA) and the EOSC Accounting System (from GRNET) as possibilities for use. 
• Based on Gergely's initial presentation on metrics (see slides attached to agenda) the following main remark was given: besides the dynamic metrics (such as energy consumption in a given period) there are static metrics that also need to be stored (such as 'energy efficiency', similar to household equipment classifications from A to G). 
• The GOCDB was found as the most suitable place for static metrics, the EGI Accounting system for dynamic metrics, however these static data probably should be reported in the accounting records similarly to the HEPSPEC normalisation values (to avoid problems caused by unavilability of GOCDB). 
• They also discussed the functionality and potential improvements of the EGI accounting system, including the normalization of CPU consumption and wall clock duration in job records. This topic needs further discussion with the EGI Accounting system team.

Next steps

1. Jiri to ask around at CESNET about possible sources for obtaining static parameters about environmental/energy friendliness of different hardware types. Inform the EGI GreenDIGIT members about this. 
2. Catalin-Gergely to schedule a follow-up meeting with the GreenDIGIT consortium to discuss the specific metrics to be stored in the metrics system, covering both static and dynamic metrics. Use the outcome of todays' meeting as baseline. We will likely need multiple set of metrics each suited for specific types of workloads (e.g. jobs, VMs, containers, IoT, ...). EGI has a decent understanding of our metrics. What about the other RIs?
3. Catalin-Gergely to schedule a meeting with the EGI Accounting (STFC, CESGA) and GOCDB Teams (STFC) to discuss an initial GreenDIGIT proposal concerning (1) the extension of the GOCDB system with static information about environmental efficiency of endpoints, (2) the extension of EGI accounting records with environmental impact related information, (3) Query, reporting and visualisation possibilities about the collected metrics. A feasability and cost analysis needs to be made about the development. 

Summary

Unified Metric for Resource Efficiency

The team discussed the need for a unified metric to compare the efficiency of different sites in consuming computing resources. They considered using the EGI accounting system, but also acknowledged its limitations. The idea of using a single green rank metric was proposed, which would be a derivative of EU or other relevant metrics. The team also discussed the possibility of using the EOSC accounting system as a backup plan. The importance of accounting for individual virtual machines was highlighted, as it would simplify the aggregation process. The team agreed to explore the possibility of using the EGI accounting system for jobs and virtual machines, while considering the use of static information from GOCDB for hardware greenness. The team also discussed the need for a backup plan for providers who don't know their carbon production, suggesting the use of energy maps or historical data to substitute numbers centrally.

 

 

 

EGI Accounting System Improvements

The team discussed the functionality and potential improvements of the EGI accounting system. Kostas and Greg confirmed that the system currently aggregates and sends summarized records to the register, not individual job records. Andrei suggested a hybrid system that could collect data from different sources, including clouds and HPC slices. Kostas clarified that the EGI accounting system is better suited for this purpose, but it would need to be extended to add new metrics. The team also discussed the scalability of the EGI system and the potential cost of adding new metrics. Gergely asked about the flexibility of the system and the possibility of tweaking the reporting side. The team agreed to further investigate the cost and feasibility of extending the EGI system to support new metrics.

 

Normalizing CPU Consumption and Benchmarks

The team discussed the normalization of CPU consumption and wall clock duration in job records. They agreed that the wall duration and CPU duration are raw time, and the scaling factor is captured in the service level fields. They also discussed the possibility of adding new benchmarks, such as carbon production, to the existing framework. The team agreed that the normalization should be done on a per-job basis, considering the efficiency of different worker nodes. They also discussed the potential of defining benchmarks in the configuration files or in a separate database, such as GOCDB. The team concluded that the benchmark information should be statically stored in the GOCDB database, but also included in the job records for convenience and to avoid processing delays.
 

GreenDIGIT Project and Benchmark Information

Alessandro discussed the current methods of retrieving benchmark information. Catalin and Gergely discussed the implications of this for the GreenDIGIT project, with Gergely suggesting that they should consider how to extend the record format to include energy consumption or CO2 emissions. Gergely also proposed that they should consider how to account for other types of services in the project. Andrei suggested that the EGI accounting record could serve as a basis for a common accounting record for all infrastructures in the GreenDIGIT project. Jaime and Gergely discussed the potential for integrating AI model training data into the EGI system, with Jaime noting that they are currently working on measuring virtual machine power consumption. The team agreed on the need to map virtual machine concepts between different systems and to consider how to integrate data from different sources into the EGI system.
 

EGI System Resolution and Metrics

In the meeting, Gergely proposed a plan to resolve the issue of appending records to the EGI system. He suggested having a follow-up meeting with a few sites, including IFCA, CNRS, CESNET, and CNRS, to discuss the metrics to be added to the records. After this, another meeting would be held with Adrian (STFC) and Carlos (CESGA) to discuss how to pick up these extensions from the records in the Accounting system. Gergely also proposed putting the EOSC Accounting system on hold until the feasibility of the EGI system  extension is figured out. Kostas agreed with this plan. Catalin suggested investigating the possibility of adding static parameters about the site, such as maximum power dissipation and thermal design power, to the records. Jiri mentioned that there is a specification called power spec, but it is not easily accessible. He promised to investigate this further for different hardware types. Gergely agreed to take some notes and circulate them to the team.