Accelerated Transient Compressible Flow Simulations using Unstructured Tetrahedral Grids on the Intel Xeon Phi Coprocessor

  • 馬 康彬

Student thesis: Master's Thesis


Conventional transient Computational Fluid Dynamics (CFD) simulations are very computationally intensive for three dimensional flows Since CFD simulation nowadays needs a larger computational grid in order to get a more precise solution as such it needs more computational time to reach a reasonable result In order to reduce the time required for CFD computation many researchers have employed parallel computing - the use of multiple CPU cores cooperating to effectively share the workload However typical High Performance Computing (HPC) clusters used for parallel CFD are very expensive making this alternative unavailable for many small engineering companies Recent developments by Intel have resulted in the release of the Xeon Phi coprocessor - a device containing a large number of CPU cores - which can be added to a conventional computer system to increase the computational capability This research presents the development and application of a transient compressible CFD solver using an unstructured tetrahedral grid to simulate three dimensional flows through complex geometries by using the computational power of the Xeon Phi coprocessor The resulting solvers - an exact Riemann solver and a QDS solver - are capable of computing flows at speeds equivalent to approximately 10-15 conventional Xeon CPU cores while only costing approximately 1/5th (20 percent) that of a conventional HPC workstation This research will cover the performance characteristics of the Many Integrated Core (MIC) Architecture of Xeon PHI coprocessor In addition to the application of an exact Riemann solver to Phi parallelization this research has applied the Quiet Direct Simulation (QDS) solver to unstructured parallel computation Details of the implementation are described within and results are shown for several industrial applications The performance characteristics of QDS compared to the analytical Riemann solver are described in detail
Date of Award2015 Jul 8
Original languageEnglish
SupervisorMatt-Hew Smith (Supervisor)

Cite this