On the distributed orchestration of stochastic discrete event simulations

Abstract

Discrete event simulations are a powerful technique for modeling stochastic systems with multiple components where interactions between these components are governed by the probability distribution functions associated with them. Complex discrete event simulations are often computationally intensive with long completion times. This paper describes our solution to the problem of orchestrating the execution of a stochastic, discrete event simulation where computational hot spots evolve spatially over time. Our performance benchmarks report on our ability to balance computational loads in these settings.The Unix version of NAADSM was written for a batch computing environment in which the modeler submits a job, and later receives completed outputs from the job, with no interaction in between. In contrast, Granules is a highly asynchronous framework in which computations may receive messages over their communication streams at any time. This mismatch was resolved by having the worker component, which is coded in Java, handle the asynchronous communications and act as a buffer for incoming messages. Communication between the worker component and the NAADSM executable is then accomplished synchronously, at specific points in NAADSM's run loop, via the stdin and stdout streams of the NAADSM processes. Consistent with our goal of loose coupling, this avoids the need to make deep changes in the NAADSM code (such as transforming it into a multithreaded program) just to support interaction with Granules.Partitioning the simulation by geography presented a challenge. Some interactions between farms do not have distance constraints associated with them, owing to the model's original scope of Figure 3. Interactions between the controller and the workers.

Scalable network analytics for characterization of outbreak influence in voluminous epidemiology datasets

“…A single commodity machine completes a typical NAADSM simulation run in about 20 hours . Subsequent analysis of the DTN is an iterative process that can far exceed simulation execution time.…”

Predictive analytics using statistical, learning, and ensemble methods to support real-time exploration of discrete event simulations

“…Parallel and distributed discrete event simulation has been well-studied in the literature [5,29,30]. While these simulations often have numerous parallelization opportunities, they also generally require fine-grained synchronization.…”

“…Each modification of the input parameters requires a new set of iterations to be run. Dividing the target simulation into several units and executing them in parallel is one way to improve overall execution times [5,6], but generally does not enable real-time exploration. In this work, we target real-time computational guarantees that involve providing subsecond, interactive responses to the user as simulation parameters are changed.…”