Multi-Level Domain-Decomposition Strategy for Solving the Eikonal Equation with the Fast-Sweeping Method

Distance field from a surface geometry is used in several scientific algorithms and applications from computer graphics, visualization, computational fluid dynamics and more. Distance field can be calculated efficiently by solving the eikonal equation at each grid point using the fast-sweeping metho...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:IEEE transactions on parallel and distributed systems Ročník 29; číslo 10; s. 2297 - 2303
Hlavní autoři: Shrestha, Anup, Senocak, Inanc
Médium: Journal Article
Jazyk:angličtina
Vydáno: New York IEEE 01.10.2018
The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
Témata:
Accelerators
Algorithms
Computational fluid dynamics
Computer graphics
Computer memory
Convergence
CUDA
Distributed memory
Domain decomposition methods
Eikonal equation
fast-sweeping method
Geometry
Graphics processing units
Heuristic algorithms
Mathematical model
MPI
OpenACC
Parallel processing
signed distance field
Strategy
Surface geometry
Sweeping
Two dimensional displays
ISSN:1045-9219, 1558-2183
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:Distance field from a surface geometry is used in several scientific algorithms and applications from computer graphics, visualization, computational fluid dynamics and more. Distance field can be calculated efficiently by solving the eikonal equation at each grid point using the fast-sweeping method. There has been an increased interest to develop parallel algorithms for the fast-sweeping method to accelerate its computational turnaround time. Most parallel strategies have focused on shared-memory parallelism and do not readily extend to distributed-memory parallelism to handle large-scale problems. To address this issue, we propose a domain-decomposition strategy to enable distributed-memory parallelism for the fast-sweeping method on heterogeneous computing clusters with accelerators. In our strategy, we make use of the Cuthill-McKee ordering for both fine- and coarse-grain parallelism such that parallel computations proceed in direction of solution characteristics. We consider both CUDA and OpenACC implementations to compute the distance field from arbitrarily complex geometries and demonstrate parallel computations of large-scale problems that necessitate distributed-memory parallelism. We also discuss the implications of the proposed strategy for scalability of the fast-sweeping method.
Bibliografie:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:1045-9219
1558-2183
DOI:10.1109/TPDS.2018.2829869