Communication Cost Consideration for Job Scheduling on Clusters

• Main Authors: Jiao-Yun Li, 李驕芸
• Other Authors: Yi-Min Wang
• Format: Others
• Language: en_US
• Published: 2005
• Online Access: http://ndltd.ncl.edu.tw/handle/kqm9fj

碩士 === 靜宜大學 === 資訊管理學系研究所 === 92 === The price of computers is more economic, and the CPU and the network are also faster and faster. We may use economic clusters to replace expensive super computers for large applications. As the previous studies for assigning resources seldom take system topology into account, our study focuses on considering the position of resources and we will give different transmission costs. We plan to evaluate the whole performance, such as slowdown, wait time, run time and turnaround time. In this study, we propose improved algorithms for FIFO and conservative. We build a trace-driven simulator to simulate the various communication competitions in a homogeneous cluster and the simulation data is from CTC workload trace. The experimental results show that although the network is faster than previous, the communication cost cannot be ignored. When the communication competition is drastic, the whole performance goes down. We concluded that it is necessary to take both the communication cost and system topology into account for job scheduling on clusters.