Abstractions and Domain Specific Language Compilers for Weather and Climate Applications
Programming Models & Languages
Scientific Software Development
TimeWednesday, June 19th9:20am - 9:40am CEST
DescriptionWeather and Climate simulations are subject of unprecedented scientific interest due to the urgent need to design and adopt efficient climate change mitigation measures.
Increasing horizontal resolution to the order of 1km is key to reduce the high uncertainty associated with climate projections. However, the computational cost of running 1km for global models on traditional CPU supercomputers is unaffordable.
Therefore, the community must adopt and be able to efficiently exploit performance of modern accelerators and hybrid supercomputing systems. However the diversity and complexity of hardware architectures is imposing a huge challenge for large and complex models that must be ported and maintained for multiple architectures combining various parallel programming models.
Several domain-specific languages (DSLs) have been developed to address the portability problem, but they usually support specific numerical methods and impose a fixed parallel programming model that decreases the scientific productivity of the model developer.
With a holistic approach to tackle the performance portability problem of complex Fortran based models, we present new high-level domain specific programming languages that allow to express numerical methods and generate efficient parallel implementations as well code generators that allow to transform existing large and complex models in Fortran.
We evaluate and discuss the different approaches based on the GPU operational code of COSMO.