Presentation
Efficient Discontinuous Galerkin Solvers for Groundwater Problems on Modern CPU Architectures
Speaker
Event Type
Focus Session
Extreme-Scale Computing
Scientific Software Development
TimeTuesday, June 18th2:45pm - 3:15pm
LocationPanorama 3
DescriptionIn order to deliver meaningful answers to real-world problems, modern earth system simulations need to model increasingly large geographical areas at a high resolution, driving a need for high performance and scalability. Discontinuous Galerkin methods are a good fit for simulations in this field, posessing beneficial mathematical properties and being well suited for modern architectures. We present our high-performance DG approach that combines a matrix-free DG solver with an AMG-based low-level subspace correction, which achieves very high efficiency on modern CPU architectures.
Hand-coding the required high-performance kernels does not scale well to application scientists such as geophysicists, who want to focus on their research topics instead of implementation details. In order to alleviate this problem, we are developing a code-generation framework that allows application scientists to formulate their problem in a high-level DSL. We will demonstrate some of the unique optimization opportunities afforded by this approach, as we are able to apply additional problem knowledge at different stages of the code generation process, which allows us to perform optimizations that would not be possible in a general-purpose code.
Hand-coding the required high-performance kernels does not scale well to application scientists such as geophysicists, who want to focus on their research topics instead of implementation details. In order to alleviate this problem, we are developing a code-generation framework that allows application scientists to formulate their problem in a high-level DSL. We will demonstrate some of the unique optimization opportunities afforded by this approach, as we are able to apply additional problem knowledge at different stages of the code generation process, which allows us to perform optimizations that would not be possible in a general-purpose code.
Speaker