Distributed experimental frame for single simulation multiple scenario experiment: Architecture, protocol definition, and implementation

Abstract

Simulation-based experiments are important tools to resolve problems too difficult to solve in modern engineering. The simulation-based experiment provides insights into all aspects of system development and acquisition, and in several ways that were not previously possible. Finding a solution using simulation-based experiments requires answering "what if" questions involving thousands or millions of different scenarios with multiple replicated trials. As a result, such simulation experiments are extremely time consuming, and many studies have attempted faster data collection methods to mitigate this.In this dissertation, we propose a Single Simulation Multiple Scenario (SSMS) architecture and its instantiation, the Distributed Experimental Frame for Simulation (DEXSim). By taking a performance-centered approach, the proposed technique makes the best use of distributed hardware resources for faster data collection in a simulation domain. Accordingly, the primary contribution of this work is to describe how the simulation-based experiment environment conducts experiments using a stochastic simulator, which improves scalability and utilizes distributed hardware resources efficiently. The SSMS describes an experiment environment with multiple processing elements that perform the same simulation on multiple scenarios simultaneously. In SSMS architecture, each simulation is independent of other simulations, so that the experimental environment for SSMS can exploit almost a full data level parallelism to collecting experiment results. In order to maximize data level parallelism, we propose a two step simulation task management scheme that utilizes a static task allocation scheme at the initial phase and a dynamic task allocation during the experiment phase. Since the replicated trials of the simulation task represent the scenario setting with different random seeds, we assume that the execution times of the replicated trials are similar. Therefore, we estimate the ...