An Algorithm for Fully-reversible Optimistic Parallel Simulation

Typically, large-scale optimistic parallel simulations will spend 90% or more of the total execution time forward processing events and very little time executing rollbacks. In fact, it was recently shown that a large-scale TCP model consisting of over 1 million nodes will execute without generating any rollbacks (i.e., perfect optimistic execution is achieved). The major cost involved in forward execution is the preparation for a rollback in the form of state-saving. Using a technique called reverse computation, state-saving overheads can be greatly reduced. Here, the rollback operation is realized by executing previously processed events in reverse. However, events are retained until GVT sweeps past. In this paper, we define a new algorithm for realizing a continuum of reverse computation-based parallel simulation systems, which enables us to relax the computing of GVT and potentially further reduces the amount of memory required to execute an optimistic simulation.