Research On PADS Modeling And Performance Analysis

Based on the analysis of parallel and distributed simulation (PADS) basic algorithms and its state-of-the-art, we reason about the difficulties of its widespread application: lack of good PADS modeling methodology and tools; lack of convenient PADS performance prediction tools based on profound understanding of performance elements and they relations. Future directions on the PADS domain are proposed which include research on the existed PADS modeling defects to find out realistic method accommodating the PADS synchronization mechanisms and compatible general software design patterns based on software engineering and research on operatable PADS performance model helping modelers to predict the PADS performance at a accepted cost before implementation.MRM (Multi-Resolution Modeling) is becoming more important as large complex distributed simulations are carried out inside and out of military simulation community. Most current distributed simulations are implemented in the framework of HLA (high level architecture), which was originally defined by US DOD and has been accepted by IEEE and OMG as an industrial simulation standard. This paper analyzes the requirements of MRM in HLA -based simulations. Traditional most adopted practical method "aggregation/disaggregation" has temporal inconsistency problem because it discards the other level as it's at one level. The mainstream theoretic method "MRE (Multiple Representation Entity)" conquers this problem by propagating changes at one resolution to others, but it has weakness in mapping consistence. This paper suggested a doctrine-oriented MRM method. This novel method has good operational easiness and keeps the consistency problem simpler than MRE by enforcing unidirectional map between entities in different resolutions.Currently most parallel and distributed simulation system design employ different modeling methodology and tools in simulation modeling and simulation software system design. However, the entire computer simulation system always is a software system which the de-fact modeling language is UML. This different incurs the increased complexity of PADS construction. We propose UML as a unified languageto both simulation object modeling and PADS software system modeling. Our work focus on extending standard UML semantics to support simulation domain specific concepts including introducing virtual time in UML, using UML statechart diagram and interaction diagram with virtual time labels and constraints to describe simulation entity functional property. In addition, we create virtual time synchronization stereotypes to describe PDES time synchronization protocols, which separates the functional properties and synchronization property, thus alleviates the simulation modeling complexity and improves both simulation model reuse and synchronization reuse.Performance prediction is important to PADS potential users. A good performance model can guide design activities and conquer the complexities. Existed analytical performance models are always based on unrealistic assumptions for the purposes of mathematical tractability, such as only two processors are considered or ignorance of communication delays and rollback costs. A probalistic model is proposed for optimistic simulation based on the adherence to realistic assumptions and characterization of various overhead costs, namely non-negligible communication delay, state-saving costs, rollback costs, infrequent checkpointing and rollback costs and n processors. The performance model has been validated against the implementation results obtained on a DAWING TCI700 parallel computer and can be readily used to estimate the performance before the optimistic simulation is implemented.