Numerical Modelling Of Free-surface Flows.Devising Strategies For Parallelism.

Three programs of the free surface flows solvers family are analysed in view of parallelism. A short description of their main features is done, with emphasis on the numerical methods used: finite elements with an E.B.E. (Element By Element) implementation. The analysis of the computational effort on scalar or vector machines is then carried out. Eventually, several strategies of parallelism are tested on different machines and compared. The results show that fine grain parallelism and matrix-vector decomposition can be efficient but only with a small number of processors, and with a maximum theoretical speed-up in the range 2-4. The rate of communications between processors, especially between workstations, is clearly a bottleneck. Domain decomposition appears to be the most promising technique and a speed-up of more than 64 can be obtained on CRAY T3D with 128 processors.