Considerable research has focused on the problem of scheduling dynamically arriving independent parallel jobs on a given set of resources to improve the performance with respect to various system and user metrics. However, there has been little work on provision of Quality of Service (QoS) in space-shared parallel job scheduling, in the form of hard deadline guarantees and service differentiation. Both of these functionalities offered by system providers are very desirable to the user. On the other hand, revenue maximization along with the optimal management of resources is appealing to a service provider.
This dissertation addresses these seemingly orthogonal aspects of parallel job scheduling in stages. At first a new scheme called QoPS is developed, to provide QoS in the form of response time guarantees. Essentially, QoPS implements an admission control mechanism for jobs, and provides deadline guarantees for all accepted jobs. Secondly,
a pioneer model is proposed to enable proportional service differentiation (PSD) in job scheduling. A PSD framework would basically allow proportional allocation of resources across users based on relative priorities.
In order to address the revenue issue, two different charging models are investigated, determined by the resource provider and user respectively. Since no QoS-enabled charging model is currently deployed at any supercomputer center, a new provider-determined charging model is proposed. In this context, the impact of user tolerance towards missed deadlines is studied, as well as various techniques to further improve the overall revenue. Alternatively, a user-centric and market-based revenue approach originally proposed for non-QoS scheduling is adapted for QoS-aware scheduling. Using this charging model, an extension to QoPS called DVQoPS is being developed, that considers the opportunity cost using a history-based predictive technique and thus maximizes the overall revenue while maintaining the deadline guarantees in an integrated way.