A grid scheduler maps efficiently the resources, available to a grid resource-service provider, to the applications submitted by the users. For efficient mapping, the scheduler should be able to use the information about the availability of compute-resources and communication resources, communication delays, the application and the communication latencies and prior reservation of resources, if any. Moreover the characteristics of the heterogeneous set of compute-nodes and the characteristics of the communication network must also be considered by a grid scheduler. A generalized scheduler should be able to handle a diverse set of jobs, with arbitrary inter-dependencies among processes and arbitrary communication channel delays. The Unified Grid Scheduling Algorithm (UGSA), presented in this paper, is able to use all of the above information to respond to the changing workload and environmental conditions without causing much of an overhead. UGSA is the first scheduler, which is able to take care of both the communication latencies and application latencies while mapping DAG-model of applications. It has a genetic algorithm mode, which can be used, when jobs with requirements of high processing are being mapped. The scheduler has been tested extensively.
展开▼
