A Java-based simulation manager for optimization and response surface methodology in multiple-response parallel simulation

Abstract

The paper describes a Java based system for allocating simulation trials to a set of P parallel processors for carrying out a simulation study involving direct-search optimization or response surface methodology. Unlike distributed simulation, where a simulation model is decomposed and its parts run in a parallel environment, the parallel replications approach allows a simulation model to run to completion with a unique set of input conditions. Since a simulation study typically involves executing R replications of the model at each of S sets of input conditions, the server's task in managing a parallel replications approach is to allocate the RS x simulation trials to P client processors in a manner that balances the workload on those processors. The objective is to complete the simulation study in a time interval approaching 1/P of that which would be required of a single processor operating in a purely sequential mode. Results are reported for several Silk based simulation models run in a Visual Cafe environment for Java.