A numerical method for the direct numerical simulation of incompressible wall turbulence in rectangular and cylindrical geometries is presented. The distinctive feature resides in its design being targeted towards an efficient distributed-memory parallel computing on commodity hardware. The adopted discretization is spectral in the two homogeneous directions; fourth-order accurate, compact finite-difference schemes over a variable-spacing mesh in the wall-normal direction are key to our parallel implementation. The parallel algorithm is designed in such a way as to minimize data exchange among the computing machines, and in particular to avoid taking a global transpose of the data during the pseudo-spectral evaluation of the non-linear terms. The computing machines can then be connected to each other through low-cost network devices. The code is optimized for memory requirements, which can moreover be subdivided among the computing nodes. The layout of a simple, dedicated and optimized computing system based on commodity hardware is described. The performance of the numerical method on this computing system is evaluated and compared with that of other codes described in the literature, as well as with that of the same code implementing a commonly employed strategy for the pseudo-spectral calculation. Ó 2005 Elsevier Inc. All rights reserved.
A low-cost parallel implementation of direct numerical simulation of wall turbulence
LUCHINI, Paolo;
2006-01-01
Abstract
A numerical method for the direct numerical simulation of incompressible wall turbulence in rectangular and cylindrical geometries is presented. The distinctive feature resides in its design being targeted towards an efficient distributed-memory parallel computing on commodity hardware. The adopted discretization is spectral in the two homogeneous directions; fourth-order accurate, compact finite-difference schemes over a variable-spacing mesh in the wall-normal direction are key to our parallel implementation. The parallel algorithm is designed in such a way as to minimize data exchange among the computing machines, and in particular to avoid taking a global transpose of the data during the pseudo-spectral evaluation of the non-linear terms. The computing machines can then be connected to each other through low-cost network devices. The code is optimized for memory requirements, which can moreover be subdivided among the computing nodes. The layout of a simple, dedicated and optimized computing system based on commodity hardware is described. The performance of the numerical method on this computing system is evaluated and compared with that of other codes described in the literature, as well as with that of the same code implementing a commonly employed strategy for the pseudo-spectral calculation. Ó 2005 Elsevier Inc. All rights reserved.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
