Scaling finite difference methods in large eddy simulation of jet engine noise to the petascale: numerical methods and their efficient and automated implementation

<p>Reduction of jet engine noise has recently become a new arena of competition between aircraft manufacturers. As a relatively new field of research in computational fluid dynamics (CFD), computational aeroacoustics (CAA) prediction of jet engine noise based on large eddy simulation (LES) is a robust and accurate tool that complements the existing theoretical and experimental approaches. In order to satisfy the stringent requirements of CAA on numerical accuracy, finite difference methods in LES-based jet engine noise prediction rely on the implicitly formulated compact spatial partial differentiation and spatial filtering schemes, a crucial component of which is an embedded solver for tridiagonal linear systems spatially oriented along the three coordinate directions of the computational space. Traditionally, researchers and engineers in CAA have employed manually crafted implementations of solvers including the transposition method, the multiblock method and the Schur complement method. Algorithmically, these solvers force a trade-off between numerical accuracy and parallel scalability. Programmingwise, implementing them for each of the three coordinate directions is tediously repetitive and error-prone. ^ In this study, we attempt to tackle both of these two challenges faced by researchers and engineers. We first describe an accurate and scalable tridiagonal linear system solver as a specialization of the truncated SPIKE algorithm and strategies for efficient implementation of the compact spatial partial differentiation and spatial filtering schemes. We then elaborate on two programming models tailored for composing regular grid-based numerical applications including finite difference-based LES of jet engine noise, one based on generalized elemental subroutines and the other based on functional array programming, and the accompanying code optimization and generation methodologies. Through empirical experiments, we demonstrate that truncated SPIKE-based spatial partial differentiation and spatial filtering deliver the theoretically promised optimal scalability in weak scaling conditions and can be implemented using the two programming models with performance on par with handwritten code while significantly reducing the required programming effort.</p>

Name thesis:degree_name

Doctor of Philosophy (PhD)

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Computer Science

Year

2014

Author dc:creator

• Situ, Yingchong

Contributors dc:contributor

• Zhiyuan Li
• Ananth Y. Grama
• Robert D. Skeel
• Ahmed H. Sameh

• Computer Sciences

Repository record dc:identifier

https://docs.lib.purdue.edu/open_access_dissertations/365

OAI identifier oai:identifier

oai:docs.lib.purdue.edu:open_access_dissertations-1515