Dynamic Work Distribution for PM Algorithm

Although poor for small dynamic scales, the Particle–Mesh (PM) model allows in astrophysics good insight for large dynamic scales at low computational cost. Furthermore, it is possible to employ a very high number of particles to get high mass resolution. These properties could be exploited by suitable parallelization of the algorithm. In PM there are two types of data: the particle data, i.e. position and velocity, which are stored in one–dimensional arrays of N elements, and the grid data, i.e. density and force, which are stored in three–dimensional arrays M ×M ×M in size. Since individual particles can change cell under the action of gravitational force, parallelization is a real challenge on parallel machine and must account for the machine architecture. We have implemented a dynamic work distribution through agenda parallelism. By subdividing the work in small tasks, the implementation is well balanced, scalable and efficient also for clustered particle distributions. In this contribution we describe this efficient, load balanced, parallel implementation of PM algorithm on Cray T3E at CINECA and show its performances on cosmological simulation results.