An Evaluation of the Benefits of Fine-Grained Value-Based Scheduling on General Purpose Clusters
CCGRID '10 Proceedings of the 2010 10th IEEE/ACM International Conference on Cluster, Cloud and Grid Computing
An evaluation of the benefits of fine-grained value-based scheduling on general purpose clusters
Future Generation Computer Systems
An evaluation of the benefits of fine-grained value-based scheduling on general purpose clusters
Future Generation Computer Systems
Hi-index | 0.00 |
Grid computing holds a major promise in transforming the Internet as we know it today to a platform for using and sharing a wide array of services and resources across organizational boundaries. This dissertation investigates how the introduction of economic principles in grid resource management through a market-based approach can deliver the necessary incentives for providers to share resources and for users to use these shared resources in a well-considered manner. We show how spot markets can provide an efficient distribution of access rights among users. We thereby demonstrate how both a commodity and a single-unit Vickrey auction market can set dynamic resource prices that are tailored to time-varying demand and supply levels, so as to obtain an optimal allocation of resources among the user population. In order to support more stringent QoS requirements, we also develop a futures market organization that allows for resources to be coallocated over time. This protects users from the financial exposure of having to pay for a partial fulfillment of their requirements. We present both a centralized and a decentralized futures market organization. A key characteristic of our approach is that we do not make the assumptions of workloads being infinitely parallellizable or preemptable at zero cost, but still allow for the market mechanism to establish an outcome in a tractable manner. In addition, we present a hybrid market organization that integrates spot and futures market mechanisms in order to combine low-latency resource allocation and trading for ad hoc workloads with the ability to reserve resources for workloads that require longer-term QoS guarantees. Studying the performance of resource management schemes through real-world deployment is costly and offers limited flexibility. In support for efficient simulation-based studies on market-based resource management schemes, we present the Grid Economics Simulator. The design of the simulator's discrete-event core shows how the combination of Aspect-Oriented Programming, Java annotations and the use of continuations contribute to an efficient simulation core and a clean programming model.