Solving 2D euler equations on a multi-processor network

Abstract

On the parallel computer “Giga-Cube” a timeaccurate, 2D solver for Eulerian equations was implemented. The computation is carried out on a structured grid. Different strategies for parallelizing the application were investigated. The application was parallelized by assigning each cell a virtual processor. All cells are computed virtually parallel. Instead of mapping cells to real processors, virtual processors to real processors are mapped, which decouples the numeric from the communication and provides always the same numerical results on every network size. All computations are described relatively to their neighbors. The timeaccurate solution can be accelerated by implicit residual averaging, which require special care for an efficient parallel implementation. This requires the solution of a linear equation system in each spatial direction. The program is divided into a communication and a computational part. The communication module is responsible for the data exchange and distributes the data of the whole computational domain onto the individual processor. The parallel implementation is verified by comparing its results with the results from the sequential program. The scalability is investigated up to 512 processors and the computing power is compared to other high performance systems.