Skip to main content
NASA Advanced Supercomputing (NAS) Division World-class modeling, simulation, and supercomputing in support of the NASA mission

High-Speed Reactive Flow Simulations on Exascale Systems: Applications to Rocket Combustors

AMS Seminar Series

Speaker: Albo Voci, Flow Physics and Aeroacoustics Laboratory, Stanford University

Presentation

Seminar Slide Deck (PDF-31MB)

Abstract

Through the growth of compute power, high-fidelity simulations have become increasingly accessible and gained widespread adoption in academics and industry. Understanding physical phenomena, prediction and other design stages rely heavily on CFD. The cornerstone of these numerical simulations is the efficient use of parallel, distributed algorithms, often tuned to hardware specific microarchitectures for maximum performance. To leverage regular data indexing for faster memory access and optimized layout, structured multiblock grids arise as a natural choice. Combined with curvilinear transformations, this topological framework enhances solver fidelity by allowing for precise control over mesh density in critical regions, while maintaining the flexibility to resolve complex geometries.

Multiblock grids present numerous algorithmic challenges regarding implementation. This presentation focuses on the use of collocated, high-order finite difference schemes in multiblock grids. In the presence of grid singularities, it is shown that the geometrical conservation laws, governing the numerical relation between the mesh and the metrics (surface and volume properties), are not satisfied. This leads to spurious oscillations which lead to simulation failure. The source of the errors is identified, and an alternative method for computing the metrics is presented. The numerical discretisation scheme is modified appropriately to incorporate the new metric terms in order to restore the geometric conservation laws.

The numerical framework to handle the multiblock topology relies heavily on Legion, a task-based parallel runtime. Legion is a data centric programming system, which abstracts away the work related to complex memory hierarchies and parallelism. A partitioning algorithm, tailored for Legion, is written to handle arbitrary topologies with disparate block sizes. To efficiently handle the individual blocks, a new data accessor is formulated, which is shown to improve locality and reduce memory footprint. Moving towards performance on exascale systems, we present a development approach that achieves a 50% fraction of nominal peak performance sustained over more than half of the computational duty cycle, yielding over 30% of peak performance overall. The results hold across multiple platforms, both CPU and GPU-based.

Biography

Photo of Albo Voci
Albo Voci is a Ph.D. student in Stanford University, where he works on high-performance computing and computational fluid dynamics, with a particular focus on scalable simulation frameworks for complex flow problems. His work has contributed to the development of advanced task-based solvers for structured, multiblock grids in generalized curvilinear coordinates. His research interests include parallel computing, high fidelity numerical methods, and fluid instabilities like the Richtmyer-Meshkov instability, with practical engineering interest.