Parallel TreeSPH

Abstract

We describe PTreeSPH, a gravity treecode combined with an SPH hydrodynamics code designed for massively parallel supercomputers having distributed memory. Our computational algorithm is based on the popular TreeSPH code of Hernquist & Katz (1989). PTreeSPH utilizes a domain decomposition procedure and a synchronous hypercube communication paradigm to build self-contained subvolumes of the simulation on each processor at every timestep. Computations then proceed in a manner analogous to a serial code. We use the Message Passing Interface (MPI) communications package, making our code easily portable to a variety of parallel systems. PTreeSPH uses individual smoothing lengths and timesteps, with a communication algorithm designed to minimize exchange of information while still providing all information required to accurately perform SPH computations. We have additionally incorporated cosmology, periodic boundary conditions with forces calculated using a quadrupole Ewald summation method, and radiative cooling and heating from a parameterized ionizing background following Katz, Weinberg & Hernquist (1996). The addition of other physical processes, such as star formation, is straightforward. A cosmological simulation from z=49 to z=2 with 64^3 gas particles and 64^3 dark matter particles requires ~6000 node-hours on a Cray T3D, with a communications overhead of ~10% and is load balanced to a ~90% level. When used on the new Cray T3E, this code will be capable of performing cosmological hydrodynamical simulations down to z=0 with ~2x10^6 particles, or to z=2 with ~10^7 particles, in a reasonable amount of time. Even larger simulations will be practical in situations where the matter is not highly clustered or when periodic boundaries are not required.