By /

How we can prepare for heatwaves using GreenDIGIT tool called ThrottlePoint

Jiří Sitera, Zdeněk Šustr

GreenDIGIT Project | September 2026

We’ll introduce another piece of the mosaic of tools and solutions being developed by the GreenDIGIT project. We’ll get right to the point. We developed a tool for load shifting at the lowest level, and it turned out to be an excellent safeguard for the data center. A safeguard against the risk of insufficient cooling capacity. It therefore does not serve its primary purpose of reducing carbon impact, but rather addresses climate change impacts in a different way, as a part of continuity management.

The ThrottlePoint tool use cases

ThrottlePoint is a load-shifting tool that operates at the lowest level—once a compute job is already running on a specific node. This means we cannot choose which type of load to run or where. That is the task of a scheduler sitting above the computing nodes.ThrottlePoint’s purpose is to defer the load and power consumption to a later time by dynamically adjusting CPU and GPU frequencies of a particular compute node.

The ThrottlePoint is driven by signals, it is fully configurable and you can plug in your own signal source. The main use cases can be described by the driver signals.

  • Driver: carbon intensity of electricity
    • Primary GreenDIGIT scenario, signals from Wattnet
    • Lowering carbon impact
  • Driver: electricity cost or consumption profile
    • Lowering operational costs
  • Driver: data center temperature
    • A safeguard for the data center
    • Optimize data center usage
    • Continuity of services
  • Driver: power-grid or local heating signals
    • Lowering operational costs, carbon impact

How can we prepare for the heatwaves?

During heat waves, the main weakness is that the design temperatures of cooling systems are significantly exceeded. Typical cooling systems do not have much margin for outdoor temperatures. A 100% margin, which is commonly used in electrical installations, would be unacceptably costly in the case of cooling. All it takes is a 10 degree Celsius increase in the temperature on the roof compared to previous years, and the cooling units start to fail. These are partial outages that cause a reduction in cooling capacity. Unfortunately, at such times, the IT equipment in the data center tends to increase its power consumption (for example fans in modern 1U servers can significantly increase power consumption), creating a positive feedback loop that worsens the situation. By deploying ThrottlePoint, it is possible to break this chain by reducing server power consumption in response to a detected increase in temperature in the data center. This prevents an emergency situation at the cost of reduced computing power.

During June and July 2026, the Czech national distributed infrastructure MetaCentrum had the opportunity to test the solution in practice at the data centers of several research institutions in the Czech Republic. The reduced resilience of cooling systems to high outdoor temperatures during heat waves is a concern for many facility managers. They have welcomed the idea: computing infrastructure can offer its power-regulation capabilities as a safeguard. Computing infrastructure typically occupies one-quarter to one-half of a data center’s capacity. The computing infrastructure with the ability to flexibly manage power consumption coexists in the data center alongside storage infrastructure, for which any outage would pose a major problem. As a result, hosting computing infrastructure becomes beneficial for the stability of the entire data center.

Can we better utilize our data centers?

The issue of resilience is directly linked to the data center’s utilization rate. One way to eliminate the cooling risk is to simply lower data center utilization, but that runs against what we’re trying to achieve. What can be done about it?

CESNET relies on a larger number of relatively small data centers operated by its individual members for its infrastructure. That’s our DNA: community collaboration, an academic network, and the integration of members’ computing and storage resources into a larger whole. But when you have a data center with different owners, needs, maintenance requirements, etc., how do you allocate capacity? Well, the facility manager allocates capacity (kW) and keeps a reserve. CESNET keeps a reserve when preparing a tender. The supplier keeps a reserve when preparing a proposal. As a result, the data center is underutilized.

We tested the use of ThrottlePoint also for this scenario. For the testbed we installed more computing power in the data center than it can handle under design conditions (maximum load on the computing nodes as measured by a benchmark). The reference point is the data center’s main monitoring system, where a measured temperature above a certain threshold triggers an alarm. When the temperature is 3–5 degrees Celsius below the maximum, ThrottlePoint kicks in and begins to gradually reduce the performance of the computing nodes. In this scenario, it is very important to monitor the extent to which the frequency throttling system intervenes. The goal is to populate the data center in such a way that the system intervenes as little as possible under real-world load. If it intervenes more frequently, there are already too many machines in the data center.

We conducted a proof-of-concept deployment called the ADAN testbed:

  • Located in Prague, collaboration with BIOCEV (Biotechnology and Biomedicine Centre of the Academy of Sciences and Charles University),
  • consisted of a 48-node cluster with 2x AMD EPYC 9554 (6,144 CPU cores),
  • consuming approximately 60 kW. 

As a result, running the real-world workload of MetaCentrum Grid service, we were able to improve data center utilization by about 25%. That means that in the same data center there can be 25% more computing nodes.

How does ThrottlePoint work and where it is available?

You can try out ThrottlePoint, we’d appreciate your feedback. The source code, documentation, and distribution package are available in the GreenDIGIT repository on GitHub: https://github.com/GreenDIGIT-project/throttlepoint 

At the core of its operation is the control or, more specifically, the limitation, of the maximum CPU and GPU frequencies. The tool is modular and, as described above, is designed to handle various control signals. It is intended for use in an integrated software solution with a scheduler and monitoring system, and therefore provides an API for these integrations. In general, two conditions must be met. For the system to produce control response, the relevant computing machines must be under a significant load. And the load type must be CPU/GPU intensive, not just I/O.

ThrottlePoint in bullets

  • Ability to lower power consumption of computing nodes with fast response.
  • Deployment conditions: CPU/GPU-intensive workload (not purely I/O).
  • Various control signals are supported to solve a wide range of use cases.
  • Primary integration with Wattnet to provide load shifting based on current carbon intensity of electricity.
  • Additional integration: APIs for schedulers and monitoring systems.
  • Tested in real-world conditions as a data center safeguard.

Summary

GreenDIGIT offers an integrated set of components to reduce environmental impact of RIs. This article highlights the potential of a component called ThrottlePoint, to significantly reduce the risks associated with operating data centers whose cooling technologies are not fully prepared for heat waves without increasing fragmentation in data center occupancy. Our contribution aims to make it possible to continue using older and smaller data centers, which are frequently available to universities and research organizations.