Speaker
Description
The EuroHPC ChEESE-2P Centre of Excellence (2023-2026, GA No. 101093038) is preparing 11 open-source flagship applications in Solid Earth sciences for pre-exascale and emerging high-performance computing (HPC) architectures. Alongside its activities in performance, scalability, accelerator support and portability, the project introduced basic software quality requirements as a starting point for Continuous Integration and Continuous Delivery (CI/CD) in EuroHPC systems and beyond. This contribution presents how FALL3D, a ChEESE-2P flagship code for atmospheric transport and deposition modelling, has extended this initial step by integrating continuous testing with scientific validation on HPC systems.
This contribution shows how a layered CI/CD workflow can provide continuous and traceable quality evidence while executing each check in the most appropriate environment. Reproducible GitLab CI jobs provide unit and component testing, coverage analysis, strict compiler checks, and AddressSanitizer-instrumented executions. Validation and benchmark activities requiring representative compilers, accelerators or scientific runs are instead submitted asynchronously through Slurm, requesting only the necessary CPU or GPU resources.
Outputs and restart files from six representative executions, covering three domain decompositions in both CPU and GPU configurations, are compared using predefined numerical tolerances. These comparisons detect changes that alter simulation results and inconsistencies between CPU and GPU implementations. Performance results are accepted only after scientific correctness has been verified. Benchmark executions use TALP, a lightweight monitoring tool that reports indicators based on the Performance Optimisation and Productivity (POP) methodology, including parallel efficiency, communication efficiency and load balance. A remotely triggered finalization workflow aggregates distributed states, reports, artifacts and metrics into a reproducible quality summary for each software change.
Our experience shows that code-level testing, scientific validation and performance characterization can be combined without forcing every check into a conventional CI runner or keeping runners occupied during HPC queueing and execution. The layered design separates reusable orchestration mechanisms from application-specific validation criteria, providing a pattern that can be adapted to other scientific HPC codes facing similar CI/CD constraints.
The next steps will focus on publishing and tracking coverage and TALP metrics, defining automated performance regression criteria, and incorporating energy measurements as an additional regression indicator. The scientific validation suite will also be extended with representative scenarios covering additional physical processes and previously untested code paths. Automated third-party assessment through the Software Quality Assurance as a Service (SQAaaS) platform, including verifiable software quality badges, is also planned.
Continuous delivery remains an open challenge. Security restrictions, platform-specific environments and maintenance costs limit the reuse of automated installations across EuroHPC systems. The European Environment for Scientific Software Installations (EESSI) provides a functional distribution path, and its further automation will be explored within the next ChEESE project phase (ChEESE-Factory). Overall, the FALL3D experience provides a practical path from basic quality requirements to continuous, evidence-based validation of scientific HPC software.