Return to the Table of ContentsObjective: The goal of this project is to develop numerical algorithm technology such that a single flow simulation code may run efficiently in nearly any computing environment, whether vector, cluster or massively parallel. A user need only learn a single interface to run successfully on any of the systems available to the CAS projects.
Approach: The algorithms used for the simulation of fluid flow are generally tailored to the available architecture. In a time of rapidly evolving hardware technology, this is a liability. Codes which run efficiently on a vector processor run poorly or not at all on a parallel system, so a great deal of time is spent porting software to new machines. What is needed is an algorithm which can accommodate different types of systems. The three-dimensional multi-partition method has been identified as a promising candidate for this work. On a massively parallel system, grids are sub-divided to provide work for many processors. A coarser sub-division is used for cluster applications, and no partitioning is done on vector machines. Convergence behavior and solutions are independent of the partitioning. The decision structure for partitioning grids will be designed into the flow solution code. The final link is a standardized communication protocol. The Parallel Virtual Machine (PVM) software from Oak Ridge National Laboratory, and the emerging Message Passing Interface (MPI) standard have been identified as possibilities.
Accomplishment: The multi-partition algorithm has now been demonstrated on a cluster of workstations connected by Ethernet and a highly parallel Intel Paragon. The basic (unpartitioned) solution scheme is in daily use on the NAS vector processors. In most cases excellent efficiency is demonstrated. The cluster code uses PVM message passing while the Intel code uses Intel's proprietary NX routines. The addition of an Ethernet switch to the workstation cluster further improves efficiency, and allows multiple multi-partition solvers to run simultaneously without mutual interference. This feature will be important in the final multiple-grid flow solver.
Status/Plans: The incorporation of the multi-partition algorithm into the framework of a multiple grid flow solver is underway. Both PVM and MPI will be used until superiority in performance and availability on all relevant computing platforms has been demonstrated for one or the other.
Points of Contact:
Rob F. Van der Wijngaart
NASA Ames Research Center
wijngaar@nas.nasa.gov
Return to the Table of Contents