Optimization Service: Parallelization And Distributed Computation

Optimization of complex systems is often characterized by large scale, complex constraints, high nonlinearity and multiple local optima. Lack of powerful computing resource is always a key difficulty, which prevents analyzing and solving those complex problems successfully for practical use. During the last two decades, parallel processing and parallel optimization techniques have been developed to take full advantage of computing resources from different sources, either on a mainframe computer or on network of workstations. Some effective approaches have been proposed, mostly designed for a certain class of special problems. Generalization and standardization is seldom touched upon. In most cases the scarcity of a corresponding supporting environment limits the parallel optimization scheme be easily implemented and widely used. On the other hand, it is desired that users concentrate only on describing and solving optimization problems, not on trivial maintenance for software and hardware.The main contributions of this dissertation are as follows:1) The trends of parallelization and, distributed computation for optimization field are reviewed, and some leading projects are outlined and compared, with focus on their functionality and architecture.2) A new concept of optimization service is proposed. Background and system analysis is presented. Two kinds of distributed systems, clusters and grid computing systems, are chosen as the computing systems, because of their high scalability and performance/cost ratio to other kinds of parallel computing systems.3) A cross-platform system MetaSolver is designed and prototyped to verify the idea of optimization service. Aiming at solving complex and large-scale optimization problems, MetaSolver is designed to run on clusters and grid computing systems. Multiple solvers and multiple tasks could be scheduled and executed cooperatively and interactively. Framework of MetaSolver is given and its implementation is detailed. Numerical experiments on solving Rastrigin's function and a difficult distillation column optimization problem demonstrates its effectivity.