Cooperative and decentralized workflow scheduling in global grids

Existing Grid scheduling systems, such as e-Science workflow brokers operate in tandem but lack the notion of cooperation mechanism that can lead to efficient application schedules across distributed resources. Lack of coordination exacerbates the utilization of various resources including computing cycles and network bandwidth. Moreover, current brokering systems have evolved around centralized client/server or hierarchical models. The responsibilities of the key functionalities such as resource discovery are delegated to the centralized server machines. Centralized models have well-known drawbacks regarding scalability, single point of failure, and network congestion at links leading to the server. To overcome these problems, this paper proposes a novel approach for decentralized and cooperative workflow scheduling in a dynamic and distributed Grid resource sharing environment. The participants in the system, such as the workflow brokers, resources, and users who belong to multiple control domains, work together to enable a single cooperative resource sharing environment. The proposed approach derives from a Distributed Hash Table (DHT) based d-dimensional logical index space with regard to resource discovery, coordination and overall system decentralization. The DHT-based d-dimensional index space serves as a blackboard system, where distributed participants can post and search complex coordination objects that regulate system wide scheduling decision making. With the implementation of our approach, not only the performance bottlenecks are likely to be eliminated but also efficient scheduling with enhanced scalability will be achieved. We evaluate and prove the feasibility of our approach through an extensive trace-driven simulation. In order to show the performance of the proposed approach against non-cooperative scheduling approach, we conduct experiment for different sizes of workflow. The results show that our scheduling technique can reduce the makespan up to 25% and demonstrates improved load balancing capability. We also compare the performance of the proposed approach against a centralized coordination technique and show that our approach is as efficient as the centralized technique with respect to achieving coordinated schedules.