MapReduce has become a mainstream mass data processing mode, as its crucial part, the scheduler has received extensive concerns of the industry. There are two deficiencies in the current delay scheduling algorithms. Firstly, a limitation of these policies is that all the tasks to be processed should be short as assumed, the performance of the algorithms declined serious when servers handle the tasks of different lengths. Secondly, delay scheduling algorithms based on static waiting time threshold, cannot adapt to the different user needs. To address this issue, this paper proposed a delay scheduling algorithm based on task classification. It adjusted tasks waiting time threshold dynamically according to the information of the different lengths. It shows that this algorithm outperforms previous delay scheduling algorithms in term of the job response time and load balance of the node.