基于调度历史数据在线预测作业执行时间.

Traditional runtimes based on user estimating is usually less accurate. This paper combined the categorization with the instance-based learning method, used the template similarity and numerical similarity method to find the similar jobs of the current jobs in historical data, and used historical scheduling data to predict the runtimes of the current jobs. This paper only took seven job attributes into account, which included user name, group name, queue name, application name, requested number of processors, requested runtime, requested memory. It applied genetic algorithm to train the best parameters, and used similar jobs attributes to predict runtimes. Compared with the existing method, experimental results show that the proposed prediction method achieves a similar underestimate rate on the premise of using fewer parameters, and gets a lower mean absolute error. Moreover, on the HPC2N04 and HPC2N05 datasets, the mean absolute errors reduce 43% and 77% respectively. This paper studied the effect of using online prediction to replace user estimation on job scheduling, analyzed the results and pointed out the future improvement directions. [ABSTRACT FROM AUTHOR]