| Recently, with the development of network and distributed computing technology, new demands are brought on cluster systems. The grid is composed of many heterogeneous resources. As the infrastructure of grid, to develop an efficient and practical cluster management system is necessary and imperative. Scheduler is the core part of cluster management system. A good scheduler can gather all power of the cluster, allocate resource to users efficiently and enhance the completion of jobs.This paper introduced grid, distributed system and cluster and some related issues. Then based on the existing scheduling algorithms, the paper emphatically analyzed the shortage and limitation of Min-Min and Max-Min algorithms from two aspects regarding task's equality and algorithms' performance. Based upon the multi-user situation in grids, it proposed a fair scheduling algorithm. This algorithm puts the user's priority in the primary location, and sufficiently considered the job status of every user. The primary realization of this algorithm is to dispute a quota value to every user based on the importance of the user. Then we conduct a dynamic rectification considering the user's priority, which is the ratio of users of the quota system in the cluster and the use of system resources and the allocation, in accordance with priority to real-time dynamic scheduling operations. This paper describes the factors of dynamic priority, as well as the impact of changes in the characteristics of job scheduling, and proposed calculation formula of dynamic priority. We also proofed the correctness of the formula and analyzed the complexity of fair scheduling algorithm. In the simulation, tectonic systems and users is constructed using simulated grid computing technology. We also tested the performance and scheduling of the actual fair scheduling algorithm, and compared with other algorithms. |
PDF Full Text Request