The authors describe how they parallelized a benchmark problem for parallel simulation, the Shark's World. The solution obtained is conservative, in the sense that no state information is saved and no rollbacks occur. Knowledge and exploitation of lookahead in the simulation model can lead to excellent performance. The search for lookahead in Sharks World led to a completely different solution approach. The advantages of the approach are manifold: on a serial workstation problems are solved over 20 times faster than with the usual discrete-event approach: the approach is easily parallelized and achieves high speedups. Also, excellent performance achieved by exploiting lookahead can be easily thwarted by relatively minor changes in problem specification. Any modification to the model rules that affects lookahead exploitation may require a great deal of modification to the solution approach. This fundamental problem will be suffered by any conservative synchronization method whose performance depends on lookahead.
展开▼
