Parallel programs distribute the work between many processors. The load should be distributed evenly so that some of processors are not idle (and/or some of processors are not overloaded). One of the popular techniques for the distribution is called Master-Slave Approach.
As the name implies, one of the processors is designated as master processor which has the knowledge of the overall work to be done. Therefore, the simulation is started up by the master, which spawns slaves, distributes the workload to them, and keeps control of the general scheduling.
Master-slave approaches often do not scale well with increasing numbers of CPUs since the workload of the master remains the same or even increases with increasing numbers of CPUs. For that reason, in TRANSIMS-1999 the master has nearly no tasks except initialization and synchronization. Even the output to file is done in a decentralized fashion. With the numbers of CPUs that we have tested in practice, we have never observed the master being the bottleneck of the parallelization.