A framework for quality of service aware resource management in multi-institutional Grids

A Grid is a collection of heterogeneous resources spread over multiple administrative and geographical domains. Many Grid applications require quality of service (QoS) guarantees in terms of guaranteed response time and guaranteed allocation of heterogeneous resources in multiple domains. Since QoS objectives of resource consumers in Grids are often in conflict with efficiency goals of resource providers and since Grid resources are shared by multiple applications that belong to different administrative domains, it is challenging to meet the QoS objectives of applications. Grid systems must also be robust enough to accommodate uncertainties such as those in user-estimated runtimes while meeting QoS and efficiency goals.; This thesis presents an advance reservations based middleware framework for Grids that achieves user satisfaction by providing QoS guarantees for Grid applications, cost effectiveness by efficiently utilizing resources and robustness by intelligently handling uncertain runtimes of applications. The framework provides components for each of the fabric, resource and collective layer of the Grid architecture. Within the framework, the thesis presents new application-to-resource mapping and scheduling strategies as experimental results show that traditional strategies do not give desired performance in multi-institutional environments. The thesis presents two scalable algorithms: Scaling through Subset Scheduling and Grid Scheduling with Deadlines for the NP-Complete problems of QoS constrained scheduling on non-shared and shared Grid resources, respectively. The thesis investigates different algorithms for mapping QoS constrained Grid requests to resources and presents a novel algorithm for mapping, Minimum Laxity Impact, which outperforms all other algorithms investigated in almost every respect for a wide range of workload parameters. Co-Allocation of heterogeneous resources in multiple domains is challenging but is required to meet the complex demand patterns of Grid applications. The thesis presents a novel three phase co-allocation approach, along with effective co-allocation algorithms, that outperform traditional techniques. The thesis also studies in detail the impact of error in user-estimated runtimes on system performance and presents strategies and policies to avoid substandard performance resulting from such inaccuracies.; Finally, the thesis argues that some of the resource management strategies and conclusions presented have a broader appeal and are applicable for other multi-institutional resource sharing infrastructures.