Performance study of cosmological simulations on message-passing and shared-memory multiprocessors

Abstract

In this paper we describe PKDGRAV, a parallel hierarchical tree-structured code used to conduct cosmological simulations on shared-memory and message-passing multiprocessors. We explore performance traits of cosmological N-Body simulations on 32K to 1.3 million particles, running PKDGRAV on KSR-2 and Intel Paragon multiprocessors with up to 128 nodes. We quantify the computation and communication requirements of PKDGRAV and study its scalability. We show that the shared-memory implementation performs and scales better than the message-passing. We investigate the causes of poor scalability of the Paragon implementation and identify an implementation-specific performance bottleneck in the software cache mechanism pertinent to the Paragon implementation.