collisions of particles with a solid surface; (3) determine the cell location of each particle; (4) on a statistical basis, This paper describes a new concept for the implementation of the direct simulation Monte Carlo (DSMC) method. It uses a localized compute collisions between particles occupying the same data structure based on a computational cell to achieve high perforcell; and (5) gather information on the particles residing mance, especially on workstation processors, which can also be in each cell. Although the large number of molecules in a used in parallel. Since the data structure makes it possible to freely real gas is replaced with a reduced number of model partiassign any cell to any processor, a domain decomposition can be cles, still hundreds of thousands to millions of particles found with equal calculation load on each processor while maintaining minimal communication among the nodes. Further, the new must be simulated, leading to tremendous computer power implementation strictly separates physical modeling, geometrical requirements.
issues, and organizational tasks to achieve high maintainability and
In the past it was most important to develop numerically to simplify future enhancements. Three example flow configuraefficient implementations of the DSMC algorithm, espetions are calculated with the new implementation to demonstrate its generality and performance. They include a flow through a di-cially for vector supercomputers [2,3]. However, the assoverging channel using an adapted unstructured triangulated grid, ciated overhead of building long arrays for these machines a flow around a planetary probe, and an internal flow in a contactor restricts the speedup achieved to a factor of about five used in plasma physics. The results are validated either by compariover unvectorized codes. Perhaps more significantly, while son with results obtained from other simulations or by comparison vector machines have reached a plateau in performance, with experimental data. High performance on an IBM SP2 system is achieved if problem size and number of parallel processors are workstations and parallel computers continue to see sigadapted accordingly. On 400 nodes, DSMC calculations with more nificant improvements. In addition, the vectorized codes than 100 million particles are possible.