Just-in-time cloning

In this work we focus on a new technique for making cloning of parallel simulations more efficient. Cloning provides a means for running multiple similar simulations in parallel where many computations are shared rather than repeated (Hybinette and Fujimoto, 2001). A simulation is cloned on an event for a particular set of logical processes, creating new LP (logical processes) clones. The clones diverge as messages from the new LPs arrive at uncloned LPs. Until replication, all the computations for a particular LP are shared between the clones. Simulation kernels using cloning achieve efficiency by incrementally replicating logical processes as necessary. This enables regions of the simulation that have not been affected to use the same computations for both the new and old clone. Clearly, the longer replication can be delayed, the more efficient the simulation. We hypothesize that in many cases replication takes place before it is strictly necessary. We propose just-in-time cloning that addresses this issue by relaxing the constraints of simulation cloning to further benefit from shared computations.