Distributed scheduling and data sharing in late-binding overlays

Pull-based late-binding overlays are used in some of today's largest computational grids. Job agents are submitted to resources with the duty of retrieving real workload from a central queue at runtime. This helps overcome the problems of these complex environments: heterogeneity, imprecise status information and relatively high failure rates. In addition, the late job assignment allows dynamic adaptation to changes in grid conditions or user priorities. However, as the scale grows, the central assignment queue may become a bottleneck for the whole system. This article presents a distributed scheduling architecture for late-binding overlays, which addresses this issue by letting execution nodes build a distributed hash table and delegating job matching and assignment to them. This reduces the load on the central server and makes the system much more scalable and robust. Scalability makes fine-grained scheduling possible and enables new functionalities, like the implementation of a distributed data cache on the execution nodes, which helps alleviate the commonly congested grid storage services.