Multivariate geographic clustering on the world`s first zero price/performance parallel computer
Forrest M. Hoffman, William W. Hargrove, A. Schultz · University of North Texas Digital Library (University of North Texas) · 1998
The authors present an application of multivariate non-hierarchical statistical clustering to geographic environmental data from the 48 conterminous US in order to produce maps of regions of ecological similarity, called ecoregions. These maps represent more realistic and finer scale regionalizations than those generated by the traditional technique: an expert with a marker pen. Nine input variables thought to affect the growth of vegetation are clustered at a resolution of one square kilometer. These data represent over 7.7 million map cells in a 9-dimensional data space. Denied the funding for the construction of a Beowulf-style cluster of new PCs on which to perform this analysis, the authors built a 126-node cluster out of surplus PCs--primarily Intel 486 CPUs with a host of different motherboards and connected via 10 Mb/s ethernet--obtained at no cost from federal facilities in Oak Ridge, Tennessee. The authors describe the construction of this unique and heterogeneous cluster. Running RedHat Linux with the GNU compiles and both PVM and MPI, this cluster, aptly named the Stone SouperComputer, is the first parallel computer with a price/performance ratio of zero. After developing a serial version of the iterative statistical clustering algorithm, the authors developed a parallel version of the algorithm which uses the MPI message passing routines. The parallel algorithm uses a classical master/slave organization, performs dynamic load balancing for reasonable performance on heterogeneous clusters, and saves intermediate results for easy restarting in case of hardware failure. In addition to being run on the Stone SouperComputer, the parallel algorithm was tested on other parallel platforms without code modification. Finally, the results of the geographic clustering are presented.