Parallel simulation refers to the execution of a single simulation on multiple cooperating processors. In this setting, speedup is measured relative to the execution time of the same simulation on a single processor. Lin adopts a more holistic, and perhaps more practical, view of speedup: examining the use of multiple processors to shorten the time required to construct confidence intervals within a simulation-based experimental framework.
Referring to this method as parallel independent replicated simulation (PIRS), Lin describes an environment within which simulation replications are assigned to workstations for processing. For the experiment reported, the workstations are heterogeneous and have dynamic workload characteristics. Simulation execution times are distributed exponentially and may be correlated to the output measures obtained. Furthermore, the number of replications needed to construct a confidence interval of a specified width is not a priori known. If a predefined workstation workload threshold is exceeded during the execution of a replication, the replication must be terminated and re-executed on another (less computationally loaded) workstation.
The study provides several heuristics for the selection of the workload threshold and optimum processor allocation.
The relevance of this investigation is clear. In addition to dedicated computational environments, the typical industrial research setting has hundreds or thousands of workstations to support its employees. Often these workstations are not continuously under significant computational load, and farming computation out to unloaded workstations is often a cost-effective approach. Whereas previous studies have addressed PIRS only with regard to dedicated, identical workstations, Lin’s work considers load balancing in a heterogeneous environment. While the experimental environment Lin uses is artificial, the heuristics for identifying workload thresholds and optimum processor allocation represent a valuable, albeit early, step in the PIRS discipline.