This meeting record documents the SIEC workshop held on February 13, 2025, at the Federal Environment Agency in Berlin.

As part of a research project funded by the Federal Environment Agency, the new Server Idle Energy Coefficient (SIEC) was presented and discussed with experts from industry, research, and the public sector. The workshop was divided into a presentation part explaining the coefficient and a World Café for in-depth discussion of its properties and application possibilities. The workshop took place in person at the Federal Environment Agency in Berlin and simultaneously online. This record summarizes the key insights, discussion points, and next steps.

Workshop Objective

The first accompanying workshop as part of the project took place on February 13, 2025, at the Federal Environment Agency in Berlin. Experts were brought together to specifically address the issue of energy consumption of servers in idle state.

The main goals of the workshop focused on "energy waste" in servers. A central focus was on identifying potentials in servers, rather than merely comparing them. Discussions emphasized the importance of improving the applicability of developed solutions and clarifying which target groups the metrics are actually relevant for. Special attention was paid to Load Correlation from a practical perspective, with P-Idle also considered an important aspect.

Concrete technical aspects were also addressed: Integrating RAM measurements to capture temporary memory performance was deemed necessary, and the possibility of measuring fan performance is further being investigated by x-ion. It was also agreed that no data will be released within the project duration.

This structured approach facilitated productive dialogue and practical engagement with the challenges and solutions in the field of energy-efficient software. A matrix for the brochure will be created to demonstrate the benefits for IT, server providers, and cloud service providers. The workshop made an important contribution to promoting sustainability within the software community and underscored the need to systematically address the energy consumption of servers in idle state.

Record Introduction:

On February 13, 2025, the first accompanying workshop on server energy efficiency took place at the Federal Environment Agency in Berlin. The workshop brought together experts from industry and the private sector to deal with server energy consumption in daily operations. The focus was on "energy efficiency" in servers and identifying optimization potentials.

Both theoretical and practical aspects were covered during the workshop. This approach facilitated productive dialogue and significantly contributed to promoting sustainability in the software community. All discussed points were recorded in the protocol.

Agenda

💡

Presentation Deck:

Agenda Items 1 to 2 – Presentation Part (Approx. 1 Hour) What is the SIEC Metric?

The following topics were presented and discussed with the help of the presentation deck.

1. Presentation of the Metric by Marina Köhn

Improving server utilization offers one of the greatest potentials for energy savings in data centers (DC), especially given the increasing capacity demands due to AI. A more energy-efficient server does not necessarily consume less energy at every workload, but can handle higher workloads with the same amount of energy.

The research project builds on the method developed by Dirk Harryvan (SIC). This will be expanded to include other server components like storage and GPUs, tested in practice, and, if successful, transferred to standardization.

2. Max Schulze explained the Potentials of the new SIEC Metric for Companies and the Measurement Setup

The SIEC metric enables the identification of untapped potentials in IT infrastructure that can later be monetized. While Harryvan's original SIC metric only considered the CPU, the advancement to SIEC by Ludger Ackermann (dc-e) now includes the energy factor. Energy consumption measurement is carried out via internal API interfaces such as RAPL or IPMI, making external measuring devices unnecessary and allowing seamless integration into existing server infrastructures.

Test scenarios were developed to examine various aspects of server utilization. These include CPU usage, usage, write and read operations for memory and disk, as well as network traffic.

The measurement data from the SIEC metric trial will be used as "snapshots" for further research. Servers are observed under different load conditions, and the effects on the SIEC metric are analyzed through load scenarios.

The following questions were discussed directly on site:

How is CPU usage considered? What energy data is in the snapshots?

Response: Capturing every 15 minutes via IPMI or other available APIs (Question by Conrad Wächter)

How is the metric calculated and how can its reliability be ensured?

Response: The sampling method is based on averages. The metric is validated through utilization experiments.

2. Ludger Ackermann Presents the Derivation of the SIEC Calculation Formula

Based on Harryvan's original calculation, the formula was expanded to all servers of a DC and renamed to Server Idle Energy Coefficient (instead of Server Idle Coefficient = SIC). The metric is suitable not only for individual servers but also for determining unnecessary energy consumption in the entire DC. Further adjustments require concrete measurement data—the formula is currently a theoretical derivation.

The following points were discussed afterward:

The metric is expressed as a percentage and aims to create transparency for server utilization and energy demand. A low P-Idle value is sought. However, a certain level of non-utilization must be factored in as a backup for possible server failures in cloud environments.

Important questions and insights were complemented online:

A central theme was integrating "Suspend to RAM" (deep sleep mode) into the P_idle calculation. The project team plans to include this scenario in the upcoming load tests. Interestingly, it was also mentioned that CPU manufacturers are already working on automated deep sleep modes, as discussed at the Intel Builder Day in Berlin a few years ago.

Another important technical question came from NTT regarding the calculation of CPU utilization on servers with multiple processors. These questions demonstrate the great interest of participants in the technical details and practical implementation of the SIEC metric.

After a short coffee break, we continued.

Transition to Part 2 of the Workshop – the World Café by Max Schulze.

In the subsequent World Café, the following topics were discussed in 2 groups (one online group, one group on-site at the UBA):

• What makes a good metric?

• The practical applicability of the metric in companies

Moderators changed rooms after 25 minutes. This way, each group could discuss both topics.

Agenda Item 4: World Café (Approx. 45 Minutes)

Summary of the On-Site Discussion:

What makes a good metric?

The following ideas, thoughts, and questions arose and were actively discussed on site:

Clarifications were made, such as SIEC not being a performance efficiency metric for servers, but rather providing statements about the efficiency of server operation or servers in a DC.

The question was explored whether the metric is sufficiently robust, not subject to misinterpretation or manipulation. A sufficient answer can only be provided after the scheduled field investigations.

Participants emphasized the importance of P-Idle as an important parameter that should be transparently represented in the metric and should ideally have a low value.

Participants wish for the SIEC metric to be as understandable as the PUE. Additionally, the question was raised for whom the metric is useful and what added value companies will gain from the new metric. It turned out that this metric is particularly of interest to cloud providers, while dedicated server providers, who typically manage a large number of servers, do not see the benefit for the end customer. However, they could imagine using the metric to optimize their own business models, which could bring long-term economic benefits.

Criticism was raised that SIEC is not an efficiency metric since the minimum value represents the optimum, and the maximum value represents the greatest waste. It was suggested to reverse the efficiency metric. Not the idle, but the utilization is determined and how this correlates with the share of energy consumption. Keyword: Load correlations.

The practical applicability of the metric in the company

Another point was the challenge of communicating an "energy waste metric"; instead, a more positive formulation should be chosen. To calculate P-Idle correctly, the network and storage must also be considered in the long term. Therefore, starting from P-Max would be easier. Thus, the question of what a 100% utilization even means must be clarified.

It was suggested that it would be desirable if the metric also indicated what costs could be saved with a favorable SIEC. A cost metric is currently not planned.

It would also be advantageous if the metric could be applied not only to servers but possibly also to clusters, making it easier for dedicated server providers to use the metric later.

Finally, the question was raised of how to create a transfer into communication with customers. The idea of the metric is to create more transparency in communication by offering not only the utilization of a server but comprehensive information.

Summary of the Online Discussion:

What makes a good metric?

The discussion started by examining the power dissipation of MOS semiconductors (P_v=Q_gate * f_switch), which shows a linear trend. When measuring server power, the role of fans was extensively discussed as they have a quadratic relationship with speed. It was clarified that server fans are part of the server power demand and can be measured via IPMI or BMC, and calculated with a known characteristic curve.

And the practical applicability in companies

Two important distinctions were established for the practical applicability of the metric:

• Clear boundaries of the control volume: Facility components like doors, recirculating cooler units, and side coolers are outside the SIEC metric. This creates a clear distinction of what is included in the measurement and what is not.

• Server differentiation: It was emphasized that there should be a clear distinction between physical and virtual servers when determining the metric. This is crucial for precise and meaningful measurement.