2026 Volume 34 Pages 503-513
As high-performance computing platforms become increasingly heterogeneous, performance portability has emerged as a critical concern for large-scale scientific applications. In this study, we present a case study of porting a legacy Fortran program to C++/Kokkos. The target is the cloud microphysics kernel of Nonhydrostatic ICosahedral Atmospheric Model (NICAM), which was written in Fortran and extracted as a standalone executable. The Kokkos-based NICAM microphysics kernel was executed using both OpenMP and CUDA backends on the Flow Type I and Type II subsystems and the Miyabi-G system. Performance evaluations show that, compared with the original Fortran program running on one Core Memory Group (CMG) of the Fujitsu A64FX, the Kokkos version with the OpenMP backend achieved approximately 33% of the Fortran performance on the Flow Type I subsystem (Fujitsu A64FX), 29% on the Flow Type II subsystem (Intel Xeon Gold 6230), and 123% on the Miyabi-G system (NVIDIA Grace CPU). With the CUDA backend, the Kokkos version achieved 2.53× and 5.31× speedups relative to the same baseline on the Flow Type II subsystem (NVIDIA Tesla V100) and the Miyabi-G system (NVIDIA Hopper H100), respectively. The porting process revealed several structural and architectural challenges, including limited vectorization efficiency and the overhead of launching many small parallel loops. These findings provide valuable insights for future efforts in porting large, complex weather and climate simulation kernels to the Kokkos programming model.
