Design and Evaluation of a Data-distributed Massively Parallel Implementation of a Global Optimization Algorithm - DIRECT

Design and Evaluation of a Data-distributed Massively Parallel Implementation of a Global Search Algorithm—DIRECT Jian He (ABSTRACT) The present work aims at an efficient, portable, and robust design of a data-distributed massively parallel DIRECT, the deterministic global optimization algorithm widely used in multidisciplinary engineering design, biological science, and physical science applications. The original algorithm is modified to adapt to different problem scales and optimization (exploration vs. exploitation) goals. Enhanced with a memory reduction technique, dynamic data structures are used to organize local data, handle unpredictable memory requirements , reduce the memory usage, and share the data across multiple processors. The parallel scheme employs a multilevel functional and data parallelism to boost concurrency and mitigate the data dependency, thus improving the load balancing and scalability. In addition, checkpointing features are integrated to provide fault tolerance and hot restarts. Important algorithm modifications and design considerations are discussed regarding data structures, parallel schemes, error handling, and portability. Using several benchmark functions and real-world applications, the present work is evaluated in terms of optimization effectiveness, data structure efficiency, memory usage, parallel performance, and checkpointing overhead. Modeling and analysis techniques are used to investigate the design effectiveness and performance sensitivity under various problem structures, parallel schemes, and system settings. Theoretical and experimental results are compared for two parallel clusters with different system scale and network connectivity. An analytical bounding model is constructed to measure the load balancing performance under different schemes. Additionally, linear regression models are used to characterize two major overhead sources— interprocessor communication and processor idleness, and also applied to the isoefficiency functions in scalability analysis. For a variety of high-dimensional problems and large scale systems, the data-distributed massively parallel design has achieved reasonable performance. The results of the performance study provide guidance for efficient problem and scheme configuration. More importantly, the generalized design considerations and analysis techniques are beneficial for transforming many global search algorithms to become effective large scale parallel optimization tools. ACKNOWLEDGMENTS I am thankful that I was not alone on the long journey to finish this work. Accompanying by my side were my dear families, friends, professors, and collaborators who have given me tremendous support, encouragement, and help in the past five years. First of all, I would like to thank my advisor, Dr. Layne T. Watson for his continued support and guidance on the journey that has not always been straight and easy. His penetrative observation and critical …