Interactive parallel simulation environments.

摘要

The goal of this research is to develop mechanisms that enable the realization of interactive simulations for rapid analysis of complex systems and for real-time virtual environment applications. Contributions of this research include parallel algorithms that (1) allow effective what-if and alternative scenario analysis and decision making in time critical situations, (2) hide computational latencies in simulation environments with real-time constraints and (3) enable frequent interactive operations such as monitoring and steering in parallel simulation systems. The development and integration of these new mechanisms are presented, implemented and evaluated.; First, a parallel simulation cloning algorithm that supports rapid evaluation of several possible alternative futures is presented. Cloning supports decision making tasks such as determining whether or not to introduce flow restrictions in air traffic control systems. Applications that may benefit include gaming, strategic and tactical battle planning, and a variety of other optimization tasks.; Next, an optimistic I/O mechanism for reducing or eliminating computational latencies, such as image rendering operations in virtual environments is introduced. This mechanism uses a faster-than-real time simulation to predict and request images before they are needed. A roll-back computation enables recovery when unexpected external inputs invalidate the predictions. Evaluation on a shared memory multiprocessor demonstrates that virtual environments can benefit from optimistic computing.; Finally an efficient, asynchronous global virtual time (GVT) algorithm for shared memory multiprocessors is presented. The algorithm enables more frequent execution of irrevocable operations (such as I/O and other interactive events). The technique is more efficient than previous approaches because the algorithm does not require message acknowledgments, special GVT messages or FIFO delivery of messages. Also, it requires only a minimal number of shared variables and data structures. Performance measurements demonstrate that frequent GVT computation can be accomplished efficiently using this approach.