Workshop: 2026 International Workshop on Performance, Portability, and Productivity in HPC
Authors: Etienne Malaboeuf (CINES/CEA), Julien Bigot (CEA/MDLS), Kévin Obrejan and Virginie Grandgirard (CEA), Emily Bourne (SCITAS/EPFL), and Thomas Padioleau (CEA/MDLS)
Abstract: Hardware heterogeneity defines the exascale era, and scientific
application performance is increasingly dominated by memory access patterns
rather than operation count.
We analyze the performance profile of a portable semi-Lagrangian
mini-application extracted from the Gysela-X++ tokamak plasma simulation code.
While single-source portability guarantees correctness across architectures, we
show that per- formance portability remains elusive, even when a code is not
intentionally tied to specific hardware features (e.g., tensor cores, cache
size, etc.).
By embedding judiciously placed layout and loop nest tuning points into this
single-source C++ mini-application, simple automatic tuning of its main operator
sequence recovers lost performance at minimal productivity cost. This automated
approach achieves speedups of up to ×1.47 on GPUs and ×1.53 on CPUs over expert
manual optimization, while requiring a fraction of the development time.
Back to 2026 International Workshop on Performance, Portability, and Productivity in HPC Archive Listing
Back to Full Workshop Archive Listing