C3P Technical Report

Non-local Cluster Update Algorithms for Spin Models

Paul D. Coddington and Clive F. Baillie

July 1990

Published in Applications of Transputers 2, Proc. of the 2nd Annual World Conference on Transputer Applications, Southampton, U.K. (July 1990), eds. D.J. Pritchard and C.J. Scott, (IOS Press, Amsterdam, 1990).

© Copyright IOS Press.


Abstract

Parallel computers are ideally suited to the Monte Carlo simulation of spin models using the standard Metropolis algorithm, since it is regular and local. However local algorithms have the major drawback that near a phase transition the number of sweeps needed to generate a statistically independent spin configuration increases as the square of the lattice size. New algorithms have recently been developed which dramatically reduce this `critical slowing down' by updating clusters of spins at a time. The highly irregular and non-local nature of the clusters means that these algorithms are much more difficult to parallelize efficiently. Here we introduce some parallel algorithms for identifying and labeling clusters, which have been implemented on the Meiko Computing Surface and other MIMD machines using the Express parallel programming environment. These algorithms are also applicable to the problem of labeling connected components in image analysis.


PostScript version (gzip compressed)

PDF version