MRI:面向并行迭代的MapReduce模型

发布时间：2018-08-02 10:06

【摘要】：机器学习领域内的多数模型均需要通过迭代计算以求解其最优参数,而MapReduce模型在迭代计算中的缺陷不足导致其在迭代计算中无法得到广泛应用。为解决上述矛盾,基于MapReduce模型提出并实现了一种可用于模型参数求解的并行迭代模型MRI。MRI模型在保持Map以及Reduce阶段的基础上,新增了Iterate阶段以及相关通信协议,实现了迭代过程中模型参数的更新、分发与迭代控制;通过对MapReduce状态机进行增强,实现了节点任务的重用,避免了迭代过程中节点任务重复创建、初始化以及回收带来的性能开销;在任务节点实现了数据缓存,保障了数据的本地性,并在Map节点增加了基于内存的块缓存机制,进一步提高训练集加载效率,以提高整体迭代效率。基于梯度下降算法的实验结果表明:MRI模型在并行迭代计算方面性能优于MapReduce模型。
[Abstract]:Most models in the field of machine learning need iterative computation to solve their optimal parameters. However, the deficiency of MapReduce model in iterative computation makes it unable to be widely used in iterative computation. In order to solve the above problems, a parallel iterative model, MRI.MRI model, which can be used to solve the model parameters, is proposed and implemented based on the MapReduce model. On the basis of maintaining the Map and Reduce phases, the Iterate phase and the related communication protocols are added. The model parameters are updated, distributed and controlled in the iterative process, and the node tasks are reused by enhancing the MapReduce state machine, avoiding the repeated creation of the node tasks in the iterative process. The performance overhead brought by initialization and collection is realized in the task node to ensure the local data, and the memory-based block cache mechanism is added to the Map node to further improve the loading efficiency of the training set. In order to improve the overall iteration efficiency. The experimental results based on gradient descent algorithm show that the performance of MapReduce model is superior to that of MapReduce model in parallel iterative computation.
【作者单位】：内蒙古工业大学信息工程学院;
【基金】：国家自然科学基金(61540004) 内蒙古自治区自然科学基金(2014MS0608)
【分类号】：TP181

【相似文献】