Improving the performance of distributed discrete event simulation by exchange of conditional look‐ahead

Abstract

Distributed discrete event simulation is an important approach for enabling the modeling and analysis of the behavior of large systems. This approach presents a major problem, namely, the possible low performance due to the excessive overhead in synchronizing the distributed logical processes. To counter this, our approach to distributed discrete event simulation involves conservative synchronization and its acceleration using dynamic estimation of process-to-process look-ahead with a feedback mechanism. This mechanism allows for the estimation of a larger look-ahead, which may be invalidated and recalculated during the course of the simulation, if one of the processes obtains more detailed knowledge.In this work, we extend the dynamically estimated look-ahead, on the basis of the local state of the logical processes, by exchanging conditional look-aheads, in conjunction with the broadcast of invalidation announcements. A notable reduction in runtime in various cases is thus achieved, especially when the estimated look-ahead is stochastically too conservative. Figure 4. Pseudocode of the algorithm.shorter than LA 2 1 and LA 3 1 . Furthermore, LP 2 sends CLA to LP 1 with the condition that LP 1 will not modify the state of LP 2 and LP 3 with external events. If LP 1 sends an external event to LP 2 , an IA will be sent to itself as discussed previously.This alternative solution can be seen as a trade-off between the external LAs and internal LAs. On one side, any LA sent from LP i to any LP in the same group is limited by min¹LA i k jLP k belongs to the other groupº:Conditional look-ahead (CLA) can be applied to extend these LAs anyway, but the limit must be kept. On the other side, the LA / CLA sent from LP i to any LP in the other group could be extended by a condition. Whether this condition really helps to extend the LA depends strongly on the model and the partitioning. This is generally unpredictable due to the unknown state of LPs in the same group. But in some special situations, the interruption of some immediate communication loops is useful.