| Large-scale parallel simulation has proved to be a cost effective technique for studying and understanding complex real world systems. There is little doubt that parallel simulations of the future will encompass the use of thousands of processors as well as local and wide area networks to solve ever larger and more complex problems. As the size and complexity of these simulations grow, so do the demands on hardware, network, and software resources, causing performance bottlenecks that can limit the effectiveness of the parallel simulation. Each of these bottlenecks, in turn, is influenced by multiple factors, thus making them difficult to understand and characterized.; Until recently, improving performance involved analyzing huge amounts of complex simulation data to determine the cause and effects of these performance issues. Without the proper software tools to analyze and tune these bottlenecks, the process of performance tuning can be complicated and tedious. This dissertation presents a new methodology for visualizing the performance bottlenecks of parallel discrete event simulations to help the user better understand how these factors degrade simulation performance and help resolve their detrimental effects. The underlying theme of this new approach is that a visualization must not simply present performance data as a pretty picture, but must tell the user something intelligent about the performance of the system. This methodology for creating effective visualizations uses a systematic and scientific approach which starts with the analysis of performance issues and ends up with structured visual representations. This dissertation presents a formal model for creating effective visualizations and discusses multiple case studies in performance tuning based on various Time Warp simulation topologies. Two prototype visualization systems were created to explore the effectiveness of the visualization methodology. A preliminary visualization prototype was created to investigate and clarify the issues in visualization of optimistic distributed simulation, and used only a single simulation view. A second prototype was created which collected much more data, and allowed the user to select from a variety of views. A carefully structured and controlled experimental evaluation of the second prototype was conducted to determine its strengths and weaknesses. |
