Design and implementation of multilevel parallel optimization on the Intel teraflops

Abstract

Single-level parallel optimization approaches, those in which either the simulation code executes in parallel or the optimization algorithm invokes multiple simultaneous single-processor analyses, have been investigated previously and been shown to be effective in reducing the time required to compute optimal solutions. However, these approaches have clear performance limitations which point to the need for multiple levels of parallelism in order to achieve peak parallel performance. Managing multiple simultaneous instances of massively parallel simulations is a challenging software undertaking, especially if the implementation is to be flexible, extensible, and generalpurpose. This paper focuses on the design for multilevel parallelism as implemented within the DAKOTA iterator toolkit. Various parallel programming models are discussed, although emphasis is given to a masterslave implementation using the Message Passing Interface (MPI). A mathematical analysis is given on achieving peak efficiency in multilevel parallelism by selecting the most effective processor partitioning schemes. This analysis is verified in some computational experiments.