WG Green Computing

Supercomputers performance keeps increasing, having surpassed the exaflop threshold several years ago. However, delivering such computational performance requires millions of processing units, resulting in a significant increase in power consumption. For example, LineShine, currently the world’s fastest supercomputer, consumes approximately 42 MW, which is equivalent to the electricity consumption of around 24,000 households. Furthermore, supercomputers are often part of a broader infrastructure, commonly referred to as the “digital continuum”, which includes scientific instruments (e.g., telescopes or various types of sensors), cloud computing platforms, and other distributed resources. This broader infrastructure makes both reducing and accurately measuring energy consumption even more challenging.

The Green Computing working group focuses on two major challenges in the context of high-performance computing: energy measurement and energy reduction.

Measuring Power and Energy Consumption. Measuring the energy consumption of a running application is challenging for several reasons. First, modern hardware architectures expose a variety of hardware counters that enable energy measurements, but accessing these counters often incurs runtime overhead. In addition, the diversity of hardware vendors and the APIs required to access these counters hinders their ease of use. Finally, the granularity of these counters is often insufficient to estimate the energy consumption of individual functions within an application. Current tools typically provide measurements at intervals ranging from 1 ms on conventional CPUs to 100 ms on GPUs, making it difficult to accurately correlate energy measurements with application behavior.

Reducing Energy Consumption. Reducing the energy consumed by an application during execution is challenging because performance must remain within acceptable limits. Achieving this objective requires deciding where computations should be executed on heterogeneous architectures, as well as when and how to exploit available optimization mechanisms such as dynamic voltage and frequency scaling (DVFS), uncore frequency scaling (UFS), and power capping. This calls for the development of scheduling strategies that jointly optimize task placement and the use of frequency scaling and power-capping techniques. Moreover, reducing energy consumption must have a controlled impact on both the robustness of the application and the accuracy of the computed solution, which often has to satisfy quality requirements. It is therefore essential to develop methods and theoretical analyses that provide users with guarantees regarding robustness and numerical accuracy. In this context, mixed-precision algorithms can offer a better trade-off between performance and energy consumption. Indeed, in many applications, double precision is only required during the most numerically sensitive phases of the computation, while lower precision can be used elsewhere. This approach has already demonstrated significant potential for improving energy efficiency.

The “Green Computing” working group aims to identify the key scientific challenges related to measuring and reducing energy consumption, investigate novel approaches to address these challenges, and contribute to fostering and coordinating the scientific community working in this area.

Head: Amina Guermouche

Mailing list: c4p-gt-ecoresponsable@groupes.renater.fr (subscribe to the working group on myGDR to join the mailing list)