Research On The Time Management Technology In Parallel And Distributed Simulation Systems

With the progressive increase of the distributed simulation size and rapid development of the high performance computing technology, the parallel and distributed simulation is gradually becoming the new research focus. Time management mechanism is the one of the key technologies which determines the correctness and repeatability of the parallel and distributed simulation, furthermore it directly influences the holistic performance of the simulation system, therefore it is very important to make a thorough research on it.The dissertation first gives a comprehensive introduction of the virtual time system, defines several fundamental concepts of the diversified simulation time and simulation types; then it analyzes the requirements to meet the causality constraint, differentiates the synchronous and asynchronous logical process simulation. Whereafter we summarize the related work of current research on the time management algorithms and study them by category, compare their basic principles and implementation ideas, propose the necessary conditions that an ideal time management mechanism need to satisfy.Conservative strategy is the most widely used time management mechanism in current simulation applications, basing on the specific definition of lookahead concept, the dissertation studies the fixed lookahead, dynamic lookahead and lookahead formal description in detail. Later it analyzes the simultaneous events and zero lookahead problem in the parallel and distributed simulation, proposes the additional field strategy to solve the unending rollback cycle problem, and farther discuss the impact of lookahead on the PDES and HLA simulation system.Most of the ideas and concepts of TW optimistic time management mechanism continue to be used and referenced by current various algorithms. We study the anti-message/rollback, transient messages determination, time-space relation snapshot and message acknowledgement mechanisms in the TW optimistic strategy, then analyze and solve the synchronization transient message problem, asynchronization simultaneous reporting problem, discuss the consistent and inconsistent cuts along with their impact on the GVT calculation. Next we research multiple GVT implementation algorithms, such as synchronous GVT, asynchronous GVT and DC-GVT algorithms, etc. Finally we construct a performance analysis model for TW system, analyze the number of processed events, average rollback length, rollback probability and its upper bound, the analysis results have instructive signification and reference value for the future research and enhancement of the TW simulation system. Hybrid strategy to some certain extent avoids the shortcomings of the excessive conservation and extremely optimism. The dissertation first analyzes the flaws of pure optimistic mechanism, then proposes a MTB hybrid algorithm which could advance time optimistically without sending anti-messages. Furthermore it improves the calculation of global virtual time and proposes an enhanced algorithm EDC-GVT which utilizes the unblocked barrier synchronization mechanism. We not only provide a performance analysis model for the MTB algorithm but also give a thoroughly theoretical analysis of the key parameter M which denotes the algorithm performance. Finally we use the PHOLD simulation model (1) to compare the performance of MTB and BTB algorithms, the experiment results indicate that MTB algorithm is better than BTB algorithm. (2) to compare the performance of conservative algorithm, TW algorithm and MTB algorithm. The experiment results indicate that under some conditions MTB algorithm has lower rollback costs and faster event process speed, which lead to better performance than TW and conservative algorithms.Adaptive strategy could dynamically adjust the system time management mechanisms according to the simulation model features and running state variations, currently it is regarded as having the most development potential and application prosperity, and is increasingly becoming the research hotspot in the time management technology. The dissertation first proposes a HTW algorithm, which combines the MTB and TW algorithms appropriately, by analyzing the key parameter Nrisk of HTW, we propose a proactive algorithm for Nrisk, basing on which an adaptive algorithm AHTW is designed and implemented. It could predict and adjust Nrisk according to the current and historical simulation status during the simulation time advancement process and obtain relative better performance. Also we construct the performance analysis model for Nrisk proactive algorithm, assess the effect of the control coefficients on the algorithm performance. We use the PHOLD simulation model to compare the performance of AHTW, MTB and TW algorithms. Experiment results show that AHTW algorithm could adjust its optimism degree dynamically according to the simulation model variations and has better adaptability and integral performance. To fully utilize the computation and memory resources, we further study the time-space consumption problem in the optimistic synchronization mechanism and present the scope of the optimal checkpoint interval, which has significant importance for checkpoint settings.It is our essential motivation and ultimate objective to implement the research productions of this dissertation into the practical environment and applications. We use cluster system as the hardware environment to devise and implement the Linux based parallel and distributed simulation platform KD-PADSE. It completely realizes the conservative, optimistic, hybrid and adaptive four types of time management mechanism and supports the incremental state saving rollback management mechanism, through federation object encapsulation and HLA proxy techniques it can connect with the RTI seamlessly. In addition it includes a graphical simulation monitor tool KD-SMT, which provides effective support for parallel federate debugging. With the KD-PADSE simulation platform, users need not deeply understand the parallel programming theory but focus their attention on the simulation object modeling, the KD-PADSE system will deal with the parallel objects distribution and execution automatically.The dissertation concentrates on the key technologies of time management in parallel and distributed simulation, make deeply research on the conservative, optimistic, hybrid and adaptive time management strategies from algorithm design and optimization, performance analysis and experiment, application research and system implementation aspects, moreover, we successfully apply our research fruits into the parallel and distributed simulation platform KD-PADSE, which achieves satisfactory achievement and has a prosperous future.