Simulation configurations are typically based upon the full factorial or fractional factorial experiment design which are well known in the performance analysis community[58]. The objective of traditional full- and fractional-factorial experiment design is to examine the end points of key parameter ranges to determine the relative importance of the different parameters (typically through a multi-dimensional linear regression fit). With network models the number of parameters combined with the possible values results in an explosion in the number of simulation experiments that must be run to perform a viable analysis of variance.
Through the use of PRC, we propose to devise a technique that speeds up the parameter-space exploration and allows a group of simulation experiments to be run faster. The approach is called double-ended experiments. Here, we propose a set of initial and finish configurations for a given network scenario. The initial condition runs at start time zero and executes until the end time divided by two. Simultaneously, we are able to execute the ``finish'' configuration experiments in the reverse direction, starting at the end time and executing until the end time divided by two. The advantage of this approach is that we are able to employ a pseudo Time Parallel scheme[77] and ``burn the candle at both ends'' while yielding four data points per start to finish run. We note here that the two middle states will differ between the two runs. This difference can be used in the overall analysis of variance.
We propose to compare this technique to standard 
factor design
and serial experiment runs based on measures of performance and linear
model accuracy.
Finally, PRC enable the modeler to create hypothetical ``end'' state scenarios and run a simulation backwards to determine how such an ``end'' state is reached. This capability has relevance beyond the parallel simulation community and impacts the general discrete-event simulation community as well.