A software system for robust and load balanced parallel computing in heterogeneous environment
Parallel Virtual Machine (PVM) is a standard software system for parallel computing on networked computers. In a heterogeneous network environment, load balancing and robustness are very important in order to increase PVM real time performance, where different machines have dissimilar performance and the network can experience machine and/or communication link failures. As a result, a task-driven approach is proposed. The new approach dynamically assigns tasks to different hosts based on task weights and CPU loads in order to maximize the usage of actual resources and minimize execution time. It also detects computer and network failures and reassigns the tasks from the failed hosts or links to the other part of the network. A load balancing and default-free software system using the proposed approach is developed for supporting PVM. Furthermore, a test-bed which simulates the heterogeneous network environment is developed for testing and evaluating the developed software system. Some research was done previously to improve the workload balance problem in PVM. One among them is query-based approach which assigns the tasks to the computers by requesting the CPU loads of the computers periodically. Experiment results shows that the developed software system using the new task-driven approach performs 23% more than the query-based approach in a heterogeneous environment.
Sampath Kumar Shamantula,
"A software system for robust and load balanced parallel computing in heterogeneous environment"
ETD Collection for Tennessee State University.