Checkpointing Optimisation to Prepare Future Exascale Plasma Turbulence Simulations
Résumé
The advent of exascale computing has revolutionized high-performance computing (HPC) and enabled unprecedented advancements in nuclear fusion research. Simulations of plasma turbulence dynamics, such as the GYSELA code, now achieve unparalleled precision and complexity. However, this progress is accompanied by significant challenges in managing the exponential growth of data generated by these simulations. Traditional input/output (I/O) methods struggle to handle the massive data volumes, heightened concurrency, and fault-tolerance requirements inherent to exascale systems. This paper investigates the I/O bottlenecks inherent in exascale computing, with a particular focus on the checkpointing mechanisms of GYSELA. These mechanisms are critical for ensuring fault tolerance and must handle several terabytes of data efficiently to avoid undermining computational performance. We analyze the current implementation of GYSELA's checkpointing mechanism managed via the PDI data interface, identifying its limitations and proposing two alternative approaches aimed at enhancing scalability and resilience. Experiments conducted on pre-exascale architectures validate the efficiency of these methods through both strong and weak scaling benchmarks. We reduced the checkpointing execution time by a factor of four, achieving near-optimal bandwidth utilisation, and we have identified implementations well-suited for exascale architectures. Our findings suggest the potential for notable performance improvements and offer insights that could help optimise I/O operations in exascale simulations.
Domaines
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|---|
| Licence |