Automated Distributed Simulation

A mis padres, que valen más que cien maestros. ABSTRACT I ABSTRACT Despite the well known advantages of distributed processing for intensive computations like simulation, frameworks often fail to exploit them. A distributed simulation is harder to develop than a sequential one, because it is necessary to interface and map activities to processors and handle the ensuing communication and synchronization problems. Very often the designer has to explicitly specify information about distribution for the framework to exploit parallelism. This dissertation presents Automated Distributed Simulation (ADS), which allows designers to forget about distribution concerns while benefiting from the advantages. This study shows abstractions that help automate the distribution of a simulation by taking maximum advantage of inherent parallelism. ADS relies on an actor formalism to provide encapsulation of component behavior and reusability. Models of computation govern component interactions by defining execution and communication mechanisms and the notion of time; thus providing semantics. Different models of computation useful for embedded systems are surveyed to discuss possibilities for distribution. Synchronous Dataflow allows for static scheduling, therefore eliminating runtime overheads and having a higher potential for efficient parallelization. Moreover it is a popular formalism suited for a large number of embedded applications, thus is chosen for the initial implementation of ADS. We present a novel execution mechanism that produces optimal periodic admissible parallel schedules allowing the dispatching mechanism to apply pipelining techniques. The implementation is described with major emphasis on distribution issues as interfacing and mapping of activities, communication and synchronization. We have chosen Ptolemy II as the implementation framework since it provides the abstractions required to achieve ADS as open source. The Ptolemy project studies heterogeneous modeling, simulation and design of concurrent real-time, embedded systems; Ptolemy II is the current software incarnation. Experiments to explore the gain for varying values of relevant parameters such as block size, number of blocks, number of iterations and topology have been designed and run. Analytical expressions for the expected results are derived and compared with the empirical results to arrive at relative overheads. The experiments show significant gains over the original implementation. The topology of a dataflow model plays an important role. However, serialization can be mitigated by using pipelined execution. Time saved in parallelization plus linear speedup in makespans provided by ADS can help tackle hitherto infeasible simulations. The implementation result of this work is publicly available as a new feature of the latest public release …