On the Feasibility of Time-Parallel Discrete Event Simulations Over Local and Wide Area Networks

Abstract

Parallel applications rarely exhibit a linear speedup to the number of processing elements used. The obstacle is the limited overlap between communication and computation. This paper discusses how the speedup of Parallel Discrete Event Simulation (PDES) is influenced when implemented on a cluster of workstations, either on a Local or a Wide Area Network. We note that optimistic space—parallel PDES algorithms (Time Warp) allow the arbitrary overlap between computation and communication, but suffer in terms of performance because of (a) the frequent communication between the Logical Processes (LPs), (b) the frequent calculation of a distributed snapshot (in the form of the Global Virtual Time (GVT)), and (c) the inherent limit up to which any model can be topologically decomposed. On the other hand, time—parallel PDES algorithms are domain specific but posses the potential for unbounded linear speedup. In order to realize this potential, a new formulation of the temporal decomposition PDES algorithm is presented which is both applicable to a large class of models and avoids the degeneracy present in previous algorithms. The same formulation provides a “clean” way to vary the computation load in order to perform load balancing over a heterogeneous network of processors. It is also illustrated how it allows the communication between the logical processes (LPs) to effectively overlap with the ongoing simulation computation of the same LPs. Finally, by way of example, a particular time—parallel simulation for an ATM statistical multiplexer is presented according to the new formulation.