Grid-Enabling an Efficient Algorithm for Demanding Global Optimization Problems in Genetic Analysis

We study the implementation on grid systems of an efficient algorithm for demanding global optimization problems. Specifically, we consider problems arising in the genetic mapping of quantitative trait loci (QTL), i.e. locations in the genome that affect a trait that is measured on a continuous scale. The scheme used in standard serial codes for QTL-analysis is easy to parallelize, and we have earlier shown that it is highly suitable for implementation on grid systems when searching for a limited number of genetic loci. However, if this type of algorithm is employed for searching for larger numbers of loci the number of evaluations of the statistical model becomes overwhelming, making the computations infeasible using any type of computer system. In this paper, we instead consider the DIRECT algorithm, which has been shown to require much fewer objective function evaluations for high-dimensional problems. By removing selected parts of the frequent global communication and synchronization in the existing parallel versions of this algorithm, we show that it is possible to retain the efficiency of the original DIRECT algorithm and perform searches for QTL models involving many loci using grid systems.