| With the increasing development of the architecture of supercomputers, and because of its outstanding price/performance ratio and scalability, availability and usability, Cluster system is becoming the hotpot and trend of the High Performance Computer Architecture and Parallel Computing research. Cluster job management system software plays very important role in a cluster system. It unified the dispersive resources in a cluster system, provides a single system view for the users, is critical to give full play to the performance of cluster and improve the rate of its utilization. It can make fully use of resources in a cluster system, effectively manage cluster system, equitably and efficiently schedule jobs. Thus, many researches show solicitude for Job Management for Cluster.Based on the former works done by others, this dissertation firstly investigates the essential issues of job management for cluster, compares some representatives of Job Management system software such as PBS, LSF, Loadleveler and Condor, and concludes an evaluation system for cluster job management system software. Then we propose a service based job management system software architecture for cluster, present its design and implementation. A service is a packed component that implements certain functions. Adopting the idea of service based design facilitates the integration of applications in the distributed environment, and enables the service based software components to be reusable and portable, make the system higher available. Moreover, we also investigate several popular job-scheduling algorithms, design and implemented a simplified First-Fit based Backfilling scheduling algorithm—RB-FIFT. The experiment demonstrates that the RB-FIFT scheduling algorithm can not only solves the problem of'starvation', but also improves the rate of system utilization. Finally, we conclude the thesis and the work done by author, and give some aspects of job management system that we should do next. |
PDF Full Text Request