Predicting Performance Impacts due to Resolution Changes in Parallel Simulations

Abstract

Multi-resolution models are often used to accelerate simulation-based analysis without significantly impacting the fidelity of the simulations. We have developed a web-enabled, component-based, multi-resolution modeling and Time Warp synchronized parallel simulation environment called WESE (Web-Enabled Simulation Environment). WESE uses a methodology called Dynamic Component Substitution (DCS) to enable abstractions or refinements to a given model. However, effectively utilizing abstractions, whether they are DCS-based or not, is a complex and time-consuming task. The complexity arises because not all abstractions improve simulation performance due to a myriad of factors related to model characteristics, synchronization protocol overheads and simulation-platform configuration. The overheads involved in identifying optimal model resolution have been exacerbating effective use of multi-resolution simulations, including our DCS-based approach. In an endeavor to minimize the time taken to identify performance impacts of resolution changes, this study proposes a DCS Performance Prediction Methodology (DCSPPM). It predicts simulation performance changes due to DCS transformations via static analysis of the model. Static analysis uses platform-specific performance characteristics of components constituting the model. DCSPPM yields quantitative estimates of performance impacts which are used by the modeler to select appropriate transformations. This article presents DCSPPM, its implementation in WESE and its empirical evaluation. The inferences drawn from the experiments prove that DCSPPM estimates have errors of less than 5% for a variety of models. Furthermore, DCSPPM executes orders of magnitude faster than corresponding shortest test simulations. Note that applicability of DCSPPM is not restricted to WESE but can be extended to other Time Warp synchronized simulators.