Job Scheduling System Key Technology For Network Computing

As a key component of the network-based computing system, job scheduling system provides efficient job scheduling strategies, integrates and makes full use of distributed computing resources to improve the system utilization.Being typical types of network-based computing systems, Cluster system and grid have become the research focus. Cluster is the mainstream architecture for high performance computing because of its low cost and good scalability. With its rapid development, Cluster system is facing the challenge of supporting diverse types of application and different administration policies. On the other hand, grid is a new-style network-based computing system. Compared to cluster system, the most distinct characteristic of Grid is that it is the kind of computing system without central administrations, omniscience and strong credits among users. Meanwhile, the construction of Grid depends on users providing their own computing resources. With the resource sharing and exchanging, grid reduces the resource redundance and provides us a non-cooperative computing platform.Ranging from cluster system to grid, we discuss and summarize the job scheduling issues in these two computing systems, which focuses on (1) the job scheduling strategy in central-admin cluster system; (2) the problem of inefficient resource use in job scheduling within multi-admin cluster system; (3) the challenge of job scheduling with incomplete job information in the non-cooperative grid platform. The main contributions in this dissertation are the followings.1. The existing cluster job scheduling algorithms can't meet the demand of guaranteeing both high performance and fairness in job scheduling. To solve this problem, RB-FIFT job scheduling algorithm, which integrates the principle of Firstfit, reservation and backfilling, is proposed in Charter 1. RB-FIFT algorithm eliminates the job starvation problem and insures the scheduling fairness with the reservation strategy. Moreover, by adopting Firstfit and backfilling principle, RB-FIFT mitigates the resource fragment and promotes the job execution. The experimental result shows that, compared to the conventional FCFS and Firstfit job scheduling algorithms, (1) providing the job executing time accurately estimated, RB-FIFT achieves substantial enhancement on several performance metrics, among which, the maximum job response time is reduced by 19.5% and the average job waiting time is reduced by 54.9% at most. (2) performance results made out by RB-FIFT algorithm doesn't increase or decrease monotonically with the inaccuracy of job execution time estimate intensifying. Compared to the value made out with the accurate estimate, the average job waiting time is reduced by a maximum of 30%.2. To solve the problem of inefficient resource use in job scheduling within multi-admin cluster system, in Charter 4, we present the dynamic-renting-based cluster job scheduling strategy. The key technology to implement this strategy is discussed and a prototype called DRJSS is also given out. Performance tests prove that the dynamic-renting-based cluster job scheduling strategy achieves better performance than the existing strategies based on the fixed-partition principle or the job forwarding principle, particularly the average job waiting time is reduced by a maximum of 1.32 times. Moreover, as the job...