Performance Benchmarking and Energy monitoring for Climate Modelling

The computational intensity of climate models makes them among the most energy-demanding applications in High-Performance Computing (HPC), resulting in significant computational costs and carbon emissions. Addressing the dual challenge of improving climate predictions —by running higher resolution, more accurate and complex models— and ensuring sustainability requires innovative tools to evaluate both computational efficiency and energy consumption across diverse HPC architectures. To address this, and in the context of the Center of Excellence in Simulation of Weather and Climate in Europe (ESiWACE), we have extended the High-Performance Climate and Weather Benchmark (HPCW) framework to incorporate a standardised set of Climate Performance Metrics for Intercomparison Projects (CPMIPs) and energy consumption monitoring.

HPCW, originally designed to maintain a set of relevant and realistic, near-operational weather forecast workloads to benchmark HPC sites, can provide insights beyond generic benchmarks like High-Performance Linpack (HPL) or High-Performance Conjugate Gradients (HPCG) by focusing on domain-specific workloads.

The inclusion of CPMIPs into HPCW brings a widely accepted set of metrics specifically tailored to the particularities of climate workflows. These metrics, already recognized by the scientific community, are key to better understanding climate model performance and allow us to keep the results from the framework relevant for research and operational runs, as well as improving our capacity for multi-model multi-platform performance comparisons.

By integrating energy monitoring, HPCW enables users to evaluate how critical computational kernels in climate models perform in terms of energy consumption. Our review of energy profiling tools across EuroHPC pre-exascale systems, including MareNostrum 5, LUMI, and Leonardo, highlights a fragmented landscape. Current tools offer varying granularity and portability, but limitations such as system configurations, administrative restrictions, and hardware compatibility often hinder their application. Low-level interfaces like Running Average Power Limit (RAPL) and Performance Application Programming Interface (PAPI) counters offer precise energy measurements but are constrained by accessibility issues.

These advancements aim to improve the allocation of climate experiments, such as those conducted for the Intergovernmental Panel on Climate Change (IPCC) in Coupled Model Intercomparison Projects (CMIPs), to the most suitable HPC resources, while also identifying architectural bottlenecks before running production experiments. Additionally, by enhancing energy consumption quantification, this work contributes to ongoing efforts to measure and reduce the carbon footprint of the climate research community. Furthermore, these analyses are expected to be particularly valuable for climate researchers, especially in the context of upcoming large-scale initiatives like CMIP7, enabling them to make informed resource requests and facilitate robust multi-platform comparisons of climate model performance which were not possible in the past. We anticipate that HPC vendors can also benefit from the outcomes of our work in optimising the systems for climate modelling workloads. By combining performance and energy metrics within a unified framework, we provide critical insights that align computational advancements with sustainability goals, ensuring efficient and environmentally conscious use of HPC resources for climate research.