Job Scheduling Based On Resource Partition Management In High Performance Cluster

With the growing demand of high performance computing, designing effective, low-price and high performance computing clusters to meet the needs of different users is an important goal of small and medium-scale high-performance computing. Through the graphical parallel programming submit interface, efficient operation scheduling strategy and resources partition management mechanism system, Bluegrid high performance computing platform is easy to use, manage and of high practical value.Efficient operation scheduling strategy and resource management mechanism is the foundation of improving the performance of HPC (High Performance Computing) Clusters. Traditional cluster job scheduling strategies focused intensely on how to balance job-scheduling among computing nodes and improve ordering equality to enhance system performance. However they paid fewer attentions to the complex and diverse requirements of users. Meanwhile complex scheduling strategy and resource control mechanism lead to laborers cluster management and poor scalability.To solve the complex and diverse user needs, this dissertation provide a method which use resources partition management mechanism to satisfied different users'demands and provide different services. Through assign different partition resources to different authorized users, all partition resources are logically independent and separated. According to the workload of partitions, the dynamic partition reorganization technique can greatly simplify the cluster management system, protect the key system operations and improve system availability.Considering different needs of users, we proposed a user-based job scheduling algorithm. The algorithm takes user rights, the availability of resources and operational factors such as weight to schedule jobs in to a whole consideration. It can improve the efficiency of scheduling, shorten the response time of the operation and meet the needs of different areas of application QoS demand, which gained considerable practical value.Test results show that the dynamic resources partition mechanisms can manage cluster computing resources effectively, improve system availability and scalability. User-based job scheduling strategy can meet different user needs and allow the system to provide demand-oriented services.