Task partition schemes for parallel processing based on petri net execution models

Abstract

Partitioning a task into modules is a necessary prerequisite to the parallel processing of the task, and also a necessary work to be done prior to solve those problems of module allocation and scheduling in multiple processors of a distributed system. The two important factors to be considered for the task partition are the amount of parallelism and the communication overhead, including performance of the system. In this thesis, with Petri net as a task model, three partition schemes are presented, based on Petri net execution models under the maximum firing rule(MFR). They are called as Lock-Step Synchronization (LSS), Partial-State Branching (PSB) and Extended PSB (XPSB). The schemes focus their attention on the following two subjects which are inevitable for executing the task in the multiprocessor system: 1) how the task can be executed with maximally possible parallelism and 2) how the modules of the task can be grouped in order to have lowest possible communication overthead with maintaining the parallelism. The maximal parallelism is exploited and maintained by applying MFR to the execution of Petri nets. However, due to the distinction of the module grouping strategies resulting from their execution models, the granularity of a module running on a processor is different from each other. Thus, the communication overhead for each partition is different. We compare the three partitions with respect to two partition overheads, defined as the synchronization overhead and the message traffic. A task partition is represented by an intermodule synchronization graph (ISG), and the partition overheads are obtained from the corresponding ISG. The overheads can be reduced by increasing the asynchronous activity or local processing activity during the task partition for parallel processing. Consequently, we show that among three partition schemes, XPSB execution model partitions a task in a manner that allows maximal parallelism with lowest partition overheads. Thei...