The Parallel k-D Tree Gravity Code


next up previous
Next: The k-D Tree Up: A Performance Study of Previous: Introduction

The Parallel k-D Tree Gravity Code

The N-Body problem addresses the gravitational evolution of astrophysical systems. It models the dynamical changes of a continuous mass distribution by approximating it with a number of particles interacting solely under the influence of gravity. The most computationally expensive part of this calculation is comprised of gravitational interactions between the particles that represent the mass distribution; these interactions must be performed for 500 to 10,000 time-steps for typical astrophysical simulations. The values of choice for N are critical for the quality of astrophysical simulation results (see Figure 1 gif). It is important to use values of N that do not compromise the validity of models for the set of physical phenomena under investigation. To this end, N should be greater than for computations aiming at determining cosmological parameters through simulation of clusters of galaxies (like the density of the Universe) [24]. Other useful N-Body simulations conducted by our group use N equal to 250,000 for Local Group runs, 1.3 million (10,000 time-steps) for a Virgo cluster of galaxies (Figure 1), and 3 million (700 time-steps) for a 100Mpc CDM Cosmological Volume.

  
Figure 1: A comparison of a CRAY simulation with 67 thousand particles, to a KSR-1 simulation with 1.3 million particles, shown at a redshift of 2. The CRAY run was performed with a vectorized tree-code, whereas the KSR run used the PKDGRAV code.

Improvements in the performance and quality of N-Body simulations has been sought in four general areas: 1) faster calculation of the gravitational accelerations; 2) multi-stepping, which is the reduction in number of time steps taken by particles in regions of the simulation where longer dynamical times are expected [24][25]; 3) volume renormalization, where regions of interest are identified and populated with a greater density of particles than the surrounding volume [21][20]; 4) the use of parallel and vector supercomputing techniques.

The need for rapid calculation of the gravitational accelerations has led to two basic approaches. The first uses grid techniques, relying mainly on the speed of FFT algorithms for the calculation of the gravitational field. This class includes the , [17] and [9] algorithms. The other approach uses multipole expansions within a hierarchical description of the mass distribution (a tree). This second class includes well known algorithms such as the Barnes-Hut tree code [2][1], the Fast Multipole Method [14][13] and other variants including the Parallel k-D Tree Gravity code (PKDGRAV).

All of the approaches mentioned above set some level of approximation in their accuracy of the calculated gravitational accelerations. Therefore, the comparison of the relative efficiency of different approaches is hard, because one has to take into account factors like the type of the physical problem under study, and acceleration-error distributions and correlations. For this reason, such a comparison is not attempted in this article. Instead, we focus on performance results for PKDGRAV that correspond to error tolerances which we have found to be adequate for typical astrophysical problems.





next up previous
Next: The k-D Tree Up: A Performance Study of Previous: Introduction