Self-Learning MapReduce Scheduler in Multi-job Environment

Abstract

Hadoop, as the most widely adopted open-source implementation of MapReduce framework, makes MapReduce widely accessible. However, it is currently limited by its default MapReduce scheduler. To achieve better performance, the scheduler should take into consideration nodes' computing power and system resources in heterogeneous environment. Further more, from job perspective, tasks' non-linear progress is also an important factor. Some research work has been carried out to enhance the performance of MapReduce, but they are not satisfactory in terms of considering characteristics of both nodes and jobs. To overcome this drawback, we propose a Self-Learning MapReduce Scheduler (SLM), which outperforms the existing schedulers in multi-job environment. Since competitions on system resources may make a task's progress unpredictable, SLM determines the progress of each job based on its own historical information. In particular, on the self-learning stage of a job, with the feedback information from the first few tasks, SLM calculates the task phase weights. With these phase weights, SLM can obtain more accurate execution time estimation, which is the most important condition to finding stragglers (slow tasks). Experimental results show that, SLM can effectively improve the accuracy of execution time estimation and straggler identification, leading to the rational utilization of resources and shortening jobs' execution time especially in multi-job environment.