gravitySimulator: The GRAPE cluster Project

A Dedicated Parallel Platform for Astrophysical Dynamics

One of the most CPU-intensive calculations in astrophysics is the gravitational N-body problem, in which a set of N particles representing stars move in response to the gravitational force generated by the N-1 other particles. The problem is hard because the number of interactions scales as O(N2), and because stars often form tightly-bound binary systems requiring short time steps. The state-of-the-art way to deal with the N-body problem is via special-purpose computers called GRAPEs (GRAvity PipEline^?). They are manufactured in Tokyo by the Hamamatsu Metrix Corporation .

The GRAPE boards compute the inverse-square force for large numbers of particles simultaneously, at speeds greatly in excess of a general-purpose supercomputer.

The next step in special-purpose hardware for the gravitational N-body problem is the GRAPE cluster ? a Beowulf cluster in which each node is connected to a GRAPE accelerator board. The largest such cluster, comprising 32 nodes, has recently been constructed at the Rochester Institute of Technology by David Merritt in the Department of Physics in collaboration with R. Spurzem of the Astronomisches Rechen-Institut at the University of Heidelberg. Called "gravitySimulator", this computational platform runs at 4 Tflops, making it one of the fastest computers in the world.

The GRAPEcluster Project is a group of students, postdocs and faculty from the astrophysics and computer science departments who meet once a week to work together on problems relating to the GRAPE cluster. The meetings are directed by Dr. Hans-Peter Bischof. Topics under discussion include algorithm development; visualization software; and scientific research, focussing on the formation and evolution of galactic nuclei and supermassive black holes. The group meets once a week in the Computer Science building. If you are interested in participating in this project, please contact David Merritt (drmsps@ritSAYNOTOSPAM.edu) or Hans-Peter Bischof (hpb@csSAYNOTOSPAM.rit.edu).

A Fortran-code, called phiGRAPE, for running N-body simulations on a GRAPE cluster is now publicly available here. The code is a parallel hermite integrator using GRAPE boards with individual block time steps. Parallelization is achieved using standard MPI communication. A performance analysis has shown that a compute speed of to 3 TFlops can be reached on the "gravitySimulator". For more details see Harfst et al. (2006).

A C-version of phiGRAPE will be available shortly.