Task assignment in parallel and distributed systems

Abstract

In this thesis, we investigate the problem of task assignment in multiple computer systems, i.e., given a set of interacting task modules comprising a distributed program to be executed on a multiple computer system, to which processor should each task module be assigned? A module is taken to be a collection of procedures of subroutines, or could be one or more data files. There are bindings or linkages between some pairs of modules since a procedure in one module may wish to transfer control to another procedure in a different module or access data contained in a different module. The task assignment problem is a fundamental aspect for distributed computing, and has received quite a lot of attention in the past decade. It arises whenever the procedures or modules of a program are distributed over several interconnected computers so that program activety moves among processors as execution proceeds. The assignment problem deals with the question of assigning task modules to processors so as to minimize the cost of running a program. The cost may be time, money or some other measure of resource usage. We consider two major problems on task assignment in multiple computer systems. The first problem is how to assign the task modules of a distributed program to the processors of a multiple processor system so as to minimize the total execution and interprocessor communication costs. The second one is how to minimize interprocessor communication costs with constraints on the degree to which the processors`` loads are balanced. The first problem is known to be NP-complete in its most general version. Previous work in this problem has shown that it is tractable for a system of two processors and for distributed programs whose task relationships are constrained. We first suggest a solution for a linear array network of any number of processors using the two-terminal network flow approach pioneered by Stone [Ston77]. We then propose an efficient modeling technique that ...