If interoperability between different simulation components is achieved usually a costly reimplementation of existing software can be avoided.
The methodology of parallel and distributed simulation is capable to achieve runtime interoperability between heterogeneous simulation components.
Further on, especially large-scale discrete event-simulations can consume a huge amount of computation time and memory resources.
In some cases, the usage of distributed and parallel simulation methods can help to reduce the overall computation time and can help to increase the scalability.
In particular the term distributed and parallel simulation refers to a spatially and concurrent execution of one simulation run.
Generally, the spatial separation is achieved by the partitioning of a simulation model in different sub-models.
Ideally, each sub-model contains a disjoint subset of the original state space whereas the set of sub-models is executed concurrently.
The coordinated or rather synchronized execution has to ensure the causality constraints.
Beside the causality issue it is often a requirement that the parallel simulation run produces the same results like a sequential one.
Events, or more closer to an implementation messages are used to exchange information between the sub-models.
The parallel or distributed execution of the spatially distributed simulation components is synchronized implicitly via events and explicitly via further synchronization mechanisms. Therefore different synchronization algortihms exists which can be separated in three classes:
- Conservative synchronization methods never break any causality constraints.
- Optimistic synchronization methods allow the execution of un-safe events but detect causality violations and restore the system state to a point before the causality violation. This procedure is called rollback.
- Hybrid synchronization methods combine advantages of both synchronization approaches.
As stated above the interprocess communication between the different sub-models or Logical Processes (LPs), a term which is used throughout the literature, is based on an exchange of events or messages respectively.
Therefore, Message Passing (MP) programming as a parallel programming model is well suited to implement distributed simulation executives.
Indeed, as underlying communication infrastructure for MatlabDS serves
the
Distributed and Parallel Application (DP) Toolbox for use with Matlab,
which provides a high-level programming interface towards PVM (Parallel Virtual Machine) and MPI-2 (Message Passing Interface). In future, a second communication platform will be supported based on the High Level Architecture. A free toolbox implementation for Matlab exists (MatlabHLA).
