Portable Parallelization of a 3-D Flow-Solver

Abstract

The parallelization of the solver reduces the coupling between the blocks of the matrices computed at each processor. The boundaries of each subdomain are exchanged with the neighbors after each relaxation step to update the boundaries for the next step. The parallel program is portable and was tested on an Intel Paragon up to 72 nodes, an IBM RS/6000 SP up to 96 nodes, a Cray T3D up to 32 nodes, a Hitachi SR2201 up to 32 nodes, and a CRAY T3E up to 512 nodes. Furthermore some runs were done on a NEC SX-4. The difference between the parallel and the sequential computed solution is negligible. The numerical efficiency is nearly the same as in the sequential program. Another very important result for this application is the scaleup because the main goal is not to compute existing problems faster but to use the code for the computation of problem sizes that could not have been computed before, because of the lack of computing power or memory. As a direct consequence of the parallelization it was possible to develop a new control mechanism for the update parameter and the CFL number. This development was only possible with the short turnaround time of the parallel program for a small case, because only this way efficient work on the problem was possible. This chapter illustrates the application of a parallel 3D simulation code for Euler- and Navier-Stokes supersonic flows around reentring space vehicles on distributed memory parallel computers. The parallelization approach uses a domain decomposition method for the subdivision of the numerical grid, the SPMD programming model and MPI. The parallel URANUS program is flexible and portable because of dynamic data structures and the MPI standard used for message passing.