Dynamic and static load balancing using PVM

This thesis presents both dynamic and static load balancing of processor work using the Parallel Virtual Machine (PVM) distributed computing environment. Im- plementations of load balancing for task queue management problems are evaluated in heterogeneous. multi-user computing environments. The objective is to use the distribution of work to the various distributed systems (nodes) and thereby mini- mize total job execution time. PVM creates processes on the nodes without regard to performance then distributes work sequentially to the processes. Static load balancing creates processes on nodes based on node performance rating plus processing load of the node during the initial assignment of work to the processes. Dynamic load balancing creates multiple processes for each node and then distributes work to processes based on node performance rating plus processing load of the node at the time the work is distributed. Static and dynamic load balancing utilize user-provided input ratings for each node and determine processing load by measure of time for execution of a standard task. Three applications are implemented using load balancing within PVM: a Mandelbrot image generator. a Conjugate-Gradient algorithm for solving linear systems of equations. and a map analysis program for landscape ecology applications. Results from the execution of these applications in a heterogeneous. multi-user. computing environment with PVM, using static and dynamic load balancing are presented. Speed improvements for dynamic and static load balancing for a map analysis habitat class are demonstrated to be as much as 2.397 and 8.830 times faster, respectively, than PVM only in an eight workstations network of five Sun4's, two IBM RS/6000 and one HP 9000/720.