A Study on Parallel Jobs Scheduling for Clusters

• Main Authors: Ko-Ta Chien, 簡克達
• Other Authors: Yi-Min Wang
• Format: Others
• Language: zh-TW
• Published: 2007
• Online Access: http://ndltd.ncl.edu.tw/handle/23436096536277903344

碩士 === 靜宜大學 === 資訊管理學系研究所 === 95 === Many scheduling policies have been developed in order to use system resources more effectively on clusters. The history of scheduling policies has progressed from the space-sharing policy such as FCFS（First-Come-First-Service）and EASY（the Extensible Argonne Scheduling sYstem）to the time-slice based and multi-processes policy --- “Gang”. The history of considering parameters for scheduling has progressed from only considering CPU resource to considering both CPU and Memory resources. The Gang scheduling with Memory consideration has proved that considering both CPU and Memory can decrease the total response time effectively. This method strictly limits system resources to provide the requirements of jobs sufficiently. It scarifies the scheduling priority of few jobs for better performance. The job might need a little more resources than previous ones, and it works. However, it takes a longer waiting time because of the constraints on Gang scheduling policy. We propose a concept about memory consideration with an upper limit. When a job doesn’t have enough resources, the scheduler will consider a greater memory size. The job may be processed after the maximum size of memory changes. The job’s waiting time is shorter, but the run time of total running jobs increases as a result of context switching. To find the balance between paging time and waiting time, in this study, we propose an improved algorithm of space-sharing policies. We build a trace-driven simulator to simulate the various memory sizes on a symmetric multiple processors clusters. Our trace data is from the DAS2 workload trace. The experimental results show that when the upper limit is 10%, the average response time of our proposed algorithm is better than some previous space-sharing policies such as FCFS and EASY.