Abstract:
Grid technology has been extensively introduced as a computing framework for aggregating the computing resources geographically distributed over the Internet. A single grid system or a single Virtual Organization can be built in the form of multiple heterogeneous computing clusters from different organizations who share the same objective. Thus, grid system has unique characteristics such as no direct communication between the computing nodes in different clusters, large data transfer overhead due to WAN latency, etc. In order to effectively use these massively computing resources within grid system, we must employ load sharing strategy to distribute workload in the system. Load sharing strategy is always one of the key components to overall performance of grid computing system. However, most strategies assign workload with respect to explicit information. This kind of information represents the characteristics of the computing resources which are difficult to be collected and unreliable to be used for making load sharing decision within grid computing system. In this work, we propose a new metric for making load decision called implicit information. It is a single metric that can represent how fast a computing node can process the submitted jobs. Moreover, it can be gathered at the coordinator node which is responsible for distributing workload during the execution. Thus, this information is comprehensive and can be used for making load decision immediately without any resource models or any monitoring services. Since implicit information cannot be used as direct substitution of explicit information, we decide to propose a new implicit strategy and its extensions for addressing unique characteristics within grid computing environment. We simulate our experiments using network simulator (NS) to evaluate the performance of our proposed strategy. We then vary the characteristics of both underlying systems and submitted applications. The obtained results of implicit strategy are compared to those from other load sharing strategy in the past. The simulation results indicate that it outperforms traditional strategies especially when information inaccuracy occurred in the system.