The cost of terminating synchronous parallel discrete-event simulations

The authors propose seven algorithms by which an arbitrary termination condition can be mechanically added to a nonterminating parallel simulation using a conservative-synchronous protocol. Informal arguments about the performance of each algorithm are made, and the arguments are confirmed through measurement of the bounded lag protocol using torus networks on a shared memory multiprocessor. Conservative-synchronous protocols are shown to be inherently efficient in the experiments considered; adding a termination algorithm requires limited additional memory and increases the wall clock time required for simulation by at most an amount proportional to the number of logical processes. Furthermore, a simulation model with any stable termination condition requires negligible overhead for the termination algorithm, a property that appears to be unique to synchronous protocols.<
>