Study of Hybrid Approaches to Task Scheduling

Modern scientific calculations may require a huge amount of computing power. Typically, such calculations are represented as a directed acyclic graphs (DAGs) that can be performed simultaneously on multiple machines in a distributed system. For large scientific problems, the amount of data and the approximate number of machine operations required for calculations are commonly known. For effective resource utilization scheduling algorithms have been created. They assign computational tasks to the machines of cluster and optimize the performance of the whole scientific application. Most planning algorithms are either static or dynamic, divided into these two groups according to their operating principle. Static methods show themselves well in the conditions of complex DAGs and accurate information about the complexity of tasks, while dynamic algorithms are able to adapt to changes in task and cluster properties. This paper proposes a hybrid approach that combines the advantages of both classes of algorithms. The approach is compared with static, dynamic and existing hybrid algorithms on synthetic and real DAGs. The experiments demonstrate the advantages of the proposed hybrid method in comparison with other approaches and demonstrate the possibility of its application for planning in real systems. In addition, the proposed approach is flexible and modular, allowing different planning strategies to be implemented. Key words: scheduling algorithms, distributed computing, hybrid scheduling methods, workflows.