A Hierarchical Parallel Scheme for a Global Search Algorithm

This paper presents a sophisticated and efficient parallel scheme for the DIRECT global optimization algorithm of Jones et al. (1993). Although several sequential implementations for this algorithm have been successfully applied to large scale MDO problems, few parallel versions of the DIRECT algorithm have addressed well algorithm characteristics such as a single starting point, an unpredictable workload, and a strong data dependency. These challenges engender many interesting design issues including domain decomposition, data access and management, and workload balancing. In the present work, a hierarchical parallel scheme has been developed to address these challenges at three levels. Each level is supported by a programming model to access shared data sets, distribute workload, or exchange messages. Results for numerical simulations and performance measurements on three test problems were obtained on a 200 node