ANSYS Benchmarking Project: Evaluation of the Domain Decomposition Solver
The focus of this paper is the Distributed Domain Solver (DDS), a prominent feature of the Parallel Performance for ANSYS (PPFA) add-on product introduced in 2001. The DDS program is based on the method of finite element tearing and interconnecting (FETI) developed by Farhat and Roux in the early ‘90s for an aerodynamic study of the high-speed civil transport aircraft.i A distinguishing attribute of the FETI algorithm is that it requires fewer interprocessor data communications than traditional domain decomposition algorithms. Convergence properties and scalability of the research code were demonstrated by Farhat, Mandel, and Roux in 1994.ii In this report, we examine resource utilization and scalability of the industrial application operating in the field. An extensible input model was created for the project, and then cases ranging from about 60,000 to over ten million degrees-of-freedom (DOF) were defined for testing purposes. A large Silicon Graphics Origin 2000 system and a commodity cluster of dual-processor Pentium III workstations were selected to perform the experiments. The Windows/NT and IRIX production release of ANSYS 5.7 were installed and testing commenced. Core sections of experimental data follow reviews of relevant system characteristics. Profiling details were fully developed for possible application within the context of the ANSYS e-CAE portal service. An effort was made to introduce key observations in proximity to the charts from which they were derived. Trailing remarks for these sections consist primarily of reflections on these points. Final comparisons of DDS scalability are made with respect to both test platforms, and a few key recommendations are presented in closing.