Parallel Implementation of Computational Fluid Dynamics Codes on Emerging Architectures

摘要

We consider two emerging parallel computing platforms and their suitability for large-scale computational mechanics codes. CRAY Multi-Threaded Architecture (MTA), featuring custom CPUs and automatic parallelizing compiler suite, is organized around flat uniform access shared memory, high-bandwidth connections between CPUs and memory, lightweight synchronization, and context switching between threads at every clock tick. On the other side of the spectrum stands the PC cluster, with commodity components and a Myrinet interconnect. This platform offers no support for automatic parallelization, necessitating the use of message-passing libraries such as MPI. Our computational code, implemented on both architectures, is based on stabilized space-time finite element formulation of incompressible Navier-Stokes equations. On the MTA, the compiler and performance analysis suite was sufficient in most cases to expose and exploit adequate parallelism in a serial version of the program. However, some data and flow structures were found to impede automatic parallelization, and needed to be modified. On the PC cluster, the bandwidth and latency of the interconnect were found sufficient for our typical applications.