Integrating multiple clusters for compute-intensive applications

Multicluster grids provide one promising solution to satisfying the growing computational demands of compute-intensive applications. However, it is challenging to seamlessly integrate all participating clusters in different domains into a single virtual computational platform. In order to fully utilize the capabilities of multicluster grids, computer scientists need to deal with the issue of joining together participating autonomic systems practically and efficiently to execute grid-enabled applications. Driven by several compute-intensive applications, this theses develops a multicluster grid management toolkit called Pelecanus to bridge the gap between user\u27s needs and the system\u27s heterogeneity. Application scientists will be able to conduct very large-scale execution across multiclusters with transparent QoS assurance. A novel model called DA-TC (Dynamic Assignment with Task Containers) is developed and is integrated into Pelecanus. This model uses the concept of a task container that allows one to decouple resource allocation from resource binding. It employs static load balancing for task container distribution and dynamic load balancing for task assignment. The slowest resources become useful rather than be bottlenecks in this manner. A cluster abstraction is implemented, which not only provides various cluster information for the DA-TC execution model, but also can be used as a standalone toolkit to monitor and evaluate the clusters\u27 functionality and performance. The performance of the proposed DA-TC model is evaluated both theoretically and experimentally. Results demonstrate the importance of reducing queuing time in decreasing the total turnaround time for an application. Experiments were conducted to understand the performance of various aspects of the DA-TC model. Experiments showed that our model could significantly reduce turnaround time and increase resource utilization for our targeted application scenarios. Four applications are implemented as case studies to determine the applicability of the DA-TC model. In each case the turnaround time is greatly reduced, which demonstrates that the DA-TC model is efficient for assisting application scientists in conducting their research. In addition, virtual resources were integrated into the DA-TC model for application execution. Experiments show that the execution model proposed in this thesis can work seamlessly with multiple hybrid grid/cloud resources to achieve reduced turnaround time