Abstract:

MapReduce models have not been widely used in iterative computation because of its defect in iterative computation. However, in order to get the optimal parameters, most of the algorithms in the field of machine learning need to be solved by iterative computation. We propose and implement a parallel iterative model based on the MapReduce for solving the optimal parameters.The MRI adds an iterate phase to the MapReduce to realize the update and distribution of parameters and the control of iteration during the iterative process. We then modify the MapReduce state machine to reuse the node tasks and avoid unnecessary performance overhead. In order to speed up the iterative process, the MRI also caches data block in the task nodes and implements the memory based block caching mechanism on the Map node. Experiment results on the gradient descent algorithm show that the performance of the proposed MRI model outperforms the MapReduce.

Key words: MapReduce, parallel computing, iterative computing, machine learning