Abstract

Pending event set (PES) structures are used in discrete event simulations to manage events according to their timestamps. A PES structure enqueues timestamped events according to priority, where the earlier events are enqueued at the head of the queue and later events at the end of the queue. The implementation of the PES structure will greatly affect its efficiency and hence the performance of discrete event simulations. This thesis looks into the performance of priority queue structures used to implement the PES. In particular, the calendar queue (CQ), which is widely accepted as having superior performance, and its variant improvements, are analyzed in detail. The analysis shows that while the CQ performs well under most situations, its performance suffers greatly when the inter-arrival pattern of events, is highly skewed and when there are large size fluctuations. Also, improvements to the CQ, such as the Dynamic Calendar Queue (DCQ) and the SNOOPy CQ continue to be plagued by costly resize operations as the queue size increases. In addition, when the queue size becomes too large, statistical variation in inter-arrival patterns becomes too generalized, making the selection of optimum operating parameters difficult. Other priority queue structures, such as the Lazy Queue and Dynamic Lazy Calendar Queue (DLCQ) make use of a `lazy' queuing algorithm to limit the primary queue size by adding another data structure to store events that are too far away for enqueuing in the primary structure. In this thesis, we propose a new lazy extension to the CQ structure, henceforth referred to as the Far future Event Leaf Tree (FELT) ii extension. FELT is based on a novel size-based algorithm that performs much better than previous lazy algorithms by effectively limiting the calend...

Citations