MapReduce framework is used for the distribution and parallelization of large-scale data processing. This framework breaks a job into several MapReduce tasks and assigns them to different nodes. A weak performance of a node in executing a task may result in a long execution of the job which is called Straggler Task. Also, detecting the nodes with the weak capability and assigning their tasks to other nodes is called Speculative Execution. This research proposes a dynamic framework to find straggler tasks in heterogeneous environments. SEWANN framework uses a neural network algorithm in order to estimate the stage weights of task execution to estimate the execution time of the tasks, accurately. Reducing the error in estimating the remaining execution time results in increasing the efficiency of big data that is the main purpose of this research. First, the proposed method was implemented in Hadoop open-source software and both estimated and actual weights were calculated. SEWANN outperformed SVR, Decision Trees, ESAMR and LATE as baseline methods 99%, 81%, 85%, and 99%, respectively. Second, SEWANN improved task execution time compared to the baseline method ESAMR by 15%, and LATE by 24%.

MapReduce框架用于大规模数据处理的分发和并行化。该框架将一项工作分解为几个MapReduce任务，并将它们分配给不同的节点。节点执行任务的性能较弱可能导致作业长时间执行，这称为Straggler任务。同样，检测能力较弱的节点并将其任务分配给其他节点也称为推测执行。这项研究提出了一个动态框架，以在异构环境中查找散乱任务。SEWANN框架使用神经网络算法来估算任务执行的阶段权重，从而准确估算任务的执行时间。减少估计剩余执行时间时的错误可提高大数据的效率，这是本研究的主要目的。首先，该方法在Hadoop开源软件中实现，并计算了估计权重和实际权重。SEWANN作为基准方法的性能分别优于SVR，决策树，ESAMR和LATE，分别为99％，81％，85％和99％。其次，与基线方法ESAMR相比，SEWANN的任务执行时间缩短了15％，而LATE则提高了24％。