The use of open-source software in high-performance computing has been crucial to furthering discovery and innovation in various fields. Recent advances in compute architectures such as Graphical Processing Units (GPUs) are enabling a new era of exascale computations. GPU devices can accomplish certain computational tasks significantly faster than CPUs can while consuming less energy. However, GPU architectures require fundamentally different programming techniques than traditional Central Processing Unit (CPU) architectures. This makes scientific software written for CPUs incompatible with GPU architectures. To address this incompatibility, developers of existing codes often choose only a single architecture or maintain multiple codes targeting different architectures. A more promising approach to address this incompatibility is to develop a universal programming paradigm for multiple compute architectures. One such technology, Kokkos, developed at Sandia National Labs, is now enabling “performance-portable” applications. That is, applications that perform well across multiple architectures. FIESTA has taken advantage of these two new technologies, GPUs and Kokkos, to provide one of the first open-source, performance-portable software applications for computational fluid dynamics.
Researchers at the University of New Mexico developed FIESTA (Fast Interface Evolution, Shocks and Transition in the Atmosphere) as a general purpose fluid-dynamics solver designed to target next generation, exascale compute clusters. FIESTA can simulate compressible fluid flows containing multiple gas species and is currently being used to study fundamental fluid physics at different scales. The problems under study include shock-accelerated interfaces, where a shockwave crosses the boundary between two gases and causes mixing; and shock-particle interactions, when a shockwave impacts a collection of solid particles suspended in a gas. FIESTA provides a modular framework and a well defined programming interface to accommodate developers with different needs and security requirements. This allows the code to be customized and used as a collaborative platform for the development of i) new methods for simulating fluid flows and ii) performance optimizations for emerging computational hardware.
- Significant improvement in execution time compared to CPU clusters
- Applicable for both GPU and CPU architectures. No duplicate codes are required.
- Particle Laden Flows
- Shock-Interface Interaction
- Forest Fire Propagation
- Atmospheric Flows
- Compressible Flows
- Turbulent Transition and Mixing
Name: Andrew Roerick