New Concepts for Virtual Testbeds : Data Mining Algorithms for Blackbox Optimization based on Wait-Free Concurrency and Generative Simulation

Abstract

Virtual testbeds have emerged as a key technology for improving and streamlining complex engineering processes by delivering long-term simulation and assessment of complex designs in virtual environments. In contrast to existing simulation technology, virtual testbeds focus on long-term physically-based simulation of the overall design in its (virtual) environment instead of only focussing on isolated, specific parts for short periods of time. This technology has the major advantage that costly testing, prototyping, and assessment in real-life environments are replaced by a cost-efficient simulation in virtual worlds for comprehensive and long-term analysis of designs. For this purpose, engineering models and their requirements are abstracted into software simulation models and objectives which are executed in virtual assessments. Simulation models are used to predict complex, real systems which can be further a subject to random influences. These predictions are used to examine the effects of individual configuration alternatives without actually realizing them and causing possible negative effects on the real system. Virtual testbeds further offer engineers the opportunity to immersively and naturally interact with their simulation model in these virtual assessments. This enables a greater and comprehensive understanding of possible design flaws early-on in the design process for engineers because they can directly assess their design in the virtual environment, based on the simulation objectives. The fact that virtual testbeds enable these realtime interactive virtual assessments, makes their underlying software infrastructure very complex. One major challenge is to minimize the development time of virtual testbeds in order to efficiently integrate them into the overall engineering process. Usually, this can be achieved by minimizing the underlying concurrency of the testbed and by simplifying its software architecture. However, this may result in a degradation of their very concurrent and asynchronous behavior, which is usually required for immersive and natural virtual interaction. A major goal of virtual testbeds in the engineering process is to find a set of optimal configurations of the simulation model which maximizes all simulation objectives for the specified virtual assessments. Once such a set has been computed, engineers can interactively explore it in the virtual environment. The main challenge is that sophisticated simulation models and their configuration are subject to a multiobjective optimization problem, which usually can not be solved manually by engineers or simulation analysts in feasible time. This is further aggravated because the relationships between simulation model configurations and simulation objectives are mostly unknown, leading to what is known as blackbox simulations. In this thesis, I propose novel data mining algorithms for computing Pareto optimal simulation model configurations, based on an approximation of the feasible design space, for deterministic and stochastic blackbox simulations in virtual testbeds for achieving above stated goal. These novel data mining algorithms lead to an automatic knowledge discovery process that does not need any supervision for its data analysis and assessment for multiobjective optimization problems of simulation model configurations. This achieves the previously stated goal of computing optimal configurations of simulation models for long-term simulations and assessments. Furthermore, I propose two complementary solutions for efficiently integrating massively-parallel virtual testbeds into engineering processes. First, I propose a novel multiversion wait-free data and concurrency management based on hash maps. These wait-free hash maps do not require any standard locking mechanisms and enable low-latency data generation, management and distribution for massively-parallel applications. Second, I propose novel concepts for efficiently code generating above wait-free data and concurrency management for arbitrary massively-parallel simulation applications of virtual testbeds. My generative simulation concept combines a state-of-the-art realtime interactive system design pattern for high maintainability with template code generation based on domain specific modelling. This concept is able to generate massively-parallel simulations and, at the same time, model checks its internal dataflow for possible interface errors. These generative concept overcomes the challenge of efficiently integrating virtual testbeds into engineering processes. These contributions enable for the first time a powerful collaboration between simulation, optimization, visualization and data analysis for novel virtual testbed applications but also overcome and achieve the presented challenges and goals.