| This work indicates that in a distributed system, load balancing is a very important part of an optimizing multi-tasks execution environment. Static and dynamic load balancing algorithms are discussed and experimental data has been acquired through the use of a production version of the VADOR framework. The Random algorithm was implemented as the representative of static load balancing methods, and the application and system characteristics are integrated into four distinct dynamic load balancing algorithms. As for dynamic load balancing algorithms, the Threshold and three versions of the Shortest (Shortest, Shortest-Capacity, Shortest-CPU) algorithms have been implemented. The Shortest, Shortest-Capacity and Shortest-CPU methods try to select a host with the lightest load value. The difference between the three methods is that they use different load indexes. The Shortest algorithm uses the CPU, memory and I/O resource load status as load index. The Shortest-Capacity algorithm uses the CPU, memory and I/O resource status and their capacities as load index. The Shortest-CPU algorithm only uses CPU load state as the load index. The Threshold algorithm use load index as in the Shortest algorithm, but also uses a Threshold state value to select a host. All dynamic load balancing algorithms use the application characteristics in the decision-making.; Experiments have been performed in different system load conditions and results have been analyzed under workloads, which mix CPU-bound, memory-bound and I/O-bound tasks together. From the experiment results, we conclude that the load balancing algorithms which use the host load status improve the system performance when compared to the Random algorithm. (Abstract shortened by UMI.)... |
