| Networks of workstations are becoming widely used to carry out resource intensive applications because of the low costs of commodity workstations and because of advances in networking technology. However, applications running on these architectures may not produce the anticipated speed-ups because of the existence of system overhead, communications congestions, and the many performance parameters that need to be tuned. This dissertation presents a new system paradigm to monitor and tune the performance of parallel applications, based on using intelligent agents to collect, analyze, and decide upon performance parameters values. This approach requires minimal effort from the application programmer and designer. This dissertation provides both an overview of this new agent-based architecture, and a detailed design for a three-level implementation.;Two case studies were developed to explore and evaluate the effectiveness of the agent-based architecture. A two-dimensional Jacobi solver was developed with the agent-based architecture attached to it on a cluster of workstations. Changes in the 2D Jacobi performance are reported and discussed to evaluate the importance of different architecture levels. A second experiment was conducted to tune the performance of a parallel Personal Communications Services (PCS) network simulator. Different simulator settings are examined and performance results were studied to determine possible enhancements to the agent-based architecture. |
