Computational Gains Using RPVM on a Beowulf Cluster

The Beowulf cluster Becker et al. (1995); Scyld ComputingCorporation (1998) is a recent advance incomputing technology that harnesses the power ofa network of desktop computers using communicationsoftware such as PVM Geist et al. (1994) and MPIMessage Passing Interface Forum (1997). Whilst thepotential of a computing cluster is obvious, expertisein programming is still developing in the statisticalcommunity.Recent articles in R-news Li and Rossini (2001)and Yu (2002) entice statistical programmers to considerwhether their solutions could be effectively calculatedin parallel. Another R package, SNOW Tierney(2002); Rossini et al. (2003) aims to skillfullyprovide a wrapper interface to these packages, independentof the underlying cluster communicationmethod used in parallel computing. This article concentrateson RPVM and wishes to build upon the contributionof Li and Rossini (2001) by taking an examplewith obvious orthogonal components and detailingthe R code necessary to allocate the computationsto each node of the cluster. The statistical techniqueused to motivate our RPVM application is the geneshavingalgorithm Hastie et al. (2000b,a) for whichS-PLUS code has been written by Do andWen (2002)to perform the calculations serially.The first section is a brief description of the Beowulfcluster used to run the R programs discussedin this paper. This is followed by an explanation ofthe gene-shaving algorithm, identifying the opportunitiesfor parallel computing of bootstrap estimatesof the "strength" of a cluster and the rendering ofeach matrix row orthogonal to the "eigen-gene". Thecode for spawning child processes is then explainedcomprehensively and the conclusion compares thespeed of RPVM on the Beowulf to serial computing.