System knowledge acquisition in parallel discrete event simulation

Abstract

Optimistic protocols designed for parallel discrete event simulation (PDES) rely heavily on the global virtual time(GVT) calculation. Since the simulation uses large amounts of memory, the GVT is used to synchronize processes and discard obsolete system information. In this paper we present a new algorithm, the continuously monitored global virtual time (CMGVT). System information, such as the local virtual time (LVT) of each process and information about messages in transit, is appended to simulation messages. We describe and analyze three variants of our GVT algorithm: direct, indirect and transitive knowledge. The direct knowledge algorithm maintains only the local information about outstanding messages. The indirect version is augmented with the information about the knowledge of its direct neighbors. Finally, the transitive version is the most comprehensive. It keeps track of the outstanding messages sent by all processes in the system.