Supercomputing researchers from ChEESE CoE develop methodology to boost code performance and portability

Researchers from Work Package 2 (WP2) of the Centre of Excellence (CoE) for Exascale in Solid Earth – ChEESE-2P project – have developed a structured methodology to improve the performance, portability, and long-term sustainability of scientific codes running on European supercomputers.

Led by the Italian Supercomputing Centre CINECA, in collaboration with the High-Performance Computing Center Stuttgart (HLRS), the Barcelona Supercomputing Center (BSC), and EVIDEN, the team has applied this approach to the project’s flagship codes and mini-apps. Now, as the project enters its final phase, the methodology is being formalised into a reusable service that can benefit the wider scientific community.

From technical audit to measurable improvement

At the core of WP2’s work are performance audits, carried out in collaboration with the Performance Optimisation and Productivity (POP) project. These audits provide a clear, quantitative picture of how efficiently a scientific code runs on large-scale supercomputers. Similar POP-based audit methodologies are also used in other European projects, including CEEC, ESiWACE, MultiXscale, SPACE, EoCoE and Plasma-PEPSC.

“Our audits assess how flagship codes perform on real scientific use-cases,” explains Piero Lanucara, Senior Technologist at CINECA and WP2 leader. “We identify the parts of the code that slow down execution – the so-called bottlenecks – and measure improvements using well-defined performance metrics.”

Each flagship code is assessed twice during the Project:

• An initial audit to establish a baseline and identify major bottlenecks
• A second audit after 24 months to measure progress and quantify improvements

Whenever possible, tests are conducted at scale on EuroHPC systems such as Leonardo, LUMI, and MareNostrum 5, ensuring realistic and comparable results.

Ensuring codes work across different supercomputers

Modern supercomputers use different hardware architectures, making it increasingly difficult to ensure that a code runs efficiently everywhere. To address this, WP2 carries out a Performance Portability Campaign using mini-apps derived from the flagship codes. The team evaluates two complementary strategies:

• Maximising performance for specific accelerators
• Prioritising portability and maintainability through unified programming models

By carefully measuring indicators such as Time to Solution and computational throughput, the team helps developers understand how to balance performance and long-term sustainability.

Preparing codes for the Exascale Era

Using the results of the audits, WP2 provides tailored optimisation recommendations for each code. Rather than applying generic solutions, the team delivers code-specific guidance that may include:

• Performance tuning
• CPU-to-GPU transitions
• Adaptation to different GPU architectures

This targeted approach helps ensure that each application is ready to take full advantage of next-generation exascale supercomputers.

Continuous monitoring for long-term sustainability

Performance optimisation is not a one-off task. To support continuous improvement, WP2 integrates TALP, a monitoring tool developed by BSC, into CI/CD pipelines. This enables automatic performance tracking, early detection of slowdowns, and faster feedback to developers – helping maintain efficiency over time.

Beyond ChEESE-2P, the WP2 methodology has been designed to be reusable and transferable. Formalised as “WP2-as-a-Service”, it includes:

• Code audits
• Mini-app extraction
• Porting and tuning support
• Performance monitoring
• Documentation and templates
• Training and dissemination

This structured workflow allows other research groups to adopt the methodology without needing to recreate the full infrastructure from scratch. “WP2 provides a rigorous and reproducible framework for HPC code optimisation,” concludes Lanucara.

According to José Gracia, head of the Scalable Programming Models and Tools Department at HLRS and Code Audit task leader for the ChEESE project, WP2’s impact extends well beyond the project itself.

“Within ChEESE, WP2 has historically acted as a catalyst for attracting community codes and channeling them – through the ChEESE first and second pase – into scientific and Grand Challenge production activities. This has enabled our community codes and expertise to contribute to projects such as DT-GEO, Geo-INQUIRE, and EuPEX, supporting advanced geoscience workflows and preparing applications for exascale computing. We expect these collaborations to continue growing, strengthening the connections between ChEESE and European HPC initiatives”, concludes Gracia.

The Methodology
It is important to note that the code audit activity described here represents one key component of a broader WP2 methodology developed within ChEESE-2P. In WP2, the support to flagship codes follows an integrated workflow that connects several steps: from the initial performance audit and analysis, to mini-app extraction in collaboration with WP3, followed by performance portability campaigns and code optimisation activities. These steps are part of an iterative process that aims to progressively improve scientific applications through successive refinements and feedback loops.

Where to read it:

DOI: https://zenodo.org/records/20327560