Rollback Reduction Techniques Through Load Balancing in Optimistic Parallel Discrete Event Simulation

Discrete event simulation is an important tool for modeling and analysis. Some of the simulation applications such as telecommunication network performance, VLSI logic circuits design, battlefield simulation, require enormous amount of computing resources. One way to satisfy this demand for computing power is to decompose the simulation system into several logical processes (Ip) and run them concurrently. In any parallel discrete event simulation (PDES) system, the events are ordered according to their time of occurrence. In order for the simulation to be correct, this ordering has to be preserved. There are three approaches to maintain this ordering. In a conservative system, no lp executes an event unless it is certain that all events with earlier time-stamps have been executed. Such systems are prone to deadlock. In an optimistic system on the other hand, simulation progresses disregarding this ordering and saves the system states regularly. Whenever a causality violation is detected, the system rolls back to a state saved earlier and restarts processing after correcting the error. There is another approach in which all the lps participate in the computation of a safe time-window and all events with time-stamps within this window are processed concurrently.\ud In optimistic simulation systems, there is a global virtual time (GVT), which is the minimum of the time-stamps of all the events existing in the system. The system can not rollback to a state prior to GVT and hence all such states can be discarded. GVT is used for memory management, load balancing, termination detection and committing of events. However, GVT computation introduces additional overhead.\ud In optimistic systems, large number of rollbacks can degrade the system performance considerably. We have studied the effect of load balancing in reducing the number of rollbacks in such systems. We have designed three load balancing algorithms and implemented two of them on a network of workstations. The other algorithm has been analyzed probabilistically. The reason for choosing network of workstations is their low cost and the availability of efficient message passing softwares like PVM and MPI. All of these load balancing algorithms piggyback on the existing GVT computation algorithms and try to balance the speed of simulation in different lps. We have also designed an optimal GVT computation algorithm for the hypercubes and studied its performance with respect to the other GVT computation algorithms by simulating a hypercube in our network cluster.\ud We use the topological properties of a star network in order to design an algorithm for computing a safe time-window for parallel discrete event simulation. We have analyzed and simulated the behavior of an open queuing network resembling such an architecture. Our algorithm is also extended for hierarchical stars and for recursive window computation