N-body Simulations on a Massively Parallel Computer

Previous abstract Next abstract

Session 6 -- Miscellaneous
Display presentation, Wednesday, January 12, 9:30-6:45, Salons I/II Room (Crystal Gateway)

[6.04] N-body Simulations on a Massively Parallel Computer

Kevin M. Olson and John E. Dorband (NASA Center for Computational Science, NASA/Goddard)

We report in this poster on progress toward an efficient algorithm for performing gravitational N-body simulations using tree data structures on a massively parallel computer architecture. We focus here on our implementation of such a tree code employing a binary tree on a Maspar MP-1 computer. The Maspar MP-1 consists of 16,384 processors, each with 64 Kbytes of memory for an aggregate of 1 Gbyte of core memory. We have run various simulations with this code ranging from simulations of interacting disk galaxies to the formation of structure in an expanding universe. We achieve speeds close to those obtained on similar problems using a Cray Y-MP. This is significant owing to the low relative cost of the Maspar MP-1. We discuss further optimizations to the basic algorithm and our progress toward achieving them. With a careful treatment of load balancing of work among the processors, we feel that speeds in excess of the fastest of today's vector supercomputers can be attained. Finally, we point out that the algorithm should scale with larger numbers of processors.

Wednesday program listing