Data Parallel Programming: Modeling of Some Hierarchic Control Structures

This paper constitutes a further step in the analysis of distributed-memory MIMD parallel machines used in dataparallel applications. In particular, this analysis considers the problem of managing the software-level interactions among various processing nodes whenever the program cannot be classified as fully data parallel because it requires some processing of data belonging to different partitions (i.e., assigned to different nodes). In previous papers, it has been shown by the authors that an efficient way of dealing with the above problem is constituted by a hierarchic control structure, decentralizing and parallelizing the management of information exchanges., Various hierarchic control structures have already been modeled in order to identify the most promising ones. In this paper, we adopt closed queuing network models to have a better insight into the expected behavior of the most effective hierarchic control structures.