A Time and Reliability Optimization Algorithm for Workflow Scheduling in Heterogeneous Distributed Computing System

Abstract

In the heterogeneous distributed computing systems, efficient task scheduling is necessary for high performance. System reliability and completion time are generally the two important metrics. In this paper, we proposed an algorithm named Merging and Duplication for Makespan and Reliability (MDMR) which combines task merging and task duplication to achieve single-objective constrained optimization of workflow reliability and completion time. The MDMR algorithm is divided into three phases: task merging, entry task duplication and reliability detection with nonentry task duplication. In the task merging phase, two tasks with a sole parent–child relationship are merged into a single task, then this process is iterated until no tasks satisfy the merging condition, thus simplifying the directed acyclic graph. In the entry task duplication phase, the entry task with the lowest reliability is duplicated on the processors, which are available at that moment to increase the reliability while reducing the workflow completion time. In the reliability detection and nonentry task duplication phase, the current workflow scheduling reliability is first calculated, then the result is directly output if the reliability is higher than the target value, otherwise the nonentry task duplication is executed with the aim of increasing the reliability until the value surpasses the target. According to the experimental results, MDMR dominates the compared algorithms and ensures that the entire workflow achieves the target reliability and shortens the workflow makespan as much as possible.