DeMS: A hybrid scheme of task scheduling and load balancing in computing clusters

In modern parallel systems and distributed applications, a large number of cores work synergistically for parallel jobs. Properly dispatching tasks among CPU cores is crucial to reduce response time of jobs, which provides benefit for both system performance and energy saving. In this paper, a hybrid scheme of task scheduling and load balancing named DeMS is proposed. DeMS consists of three algorithms, including On-Demand scheduling, Querying and Migrating Task (QMT) and Staged Task Migration (STM). The On-Demand scheduling algorithm is proposed to decrease the communication overhead between a master and slaves. Slaves have an initiative state declaring mechanism and the master can find out a slave with low workload to dispatch a new task. QTM is designed to keep the workload balanced. A slave with high workload can be detected by the master which will assign the last dispatched task to another idle slave. Besides, the dependencies among tasks are considered and STM is proposed to schedule the tasks associated with each other. A job is divided into stages according to tasks׳ execution sequence and Data Shuffling is used to represent interactions between stages. Finally, a testbed is developed to evaluate DeMS and we conduct a series of experiments on 10,000 virtual slaves. Simulation results demonstrate that our proposed On-Demand scheduling algorithm can significantly reduce the response time of parallel jobs. Meanwhile, QMT and STM are effective for independent-task and dependent-task schedulings, respectively.