Aimed at the problems that the layered parallel SVM algorithm is to generate subsample sets by completely adopting the random partition method, and the distribution deviation exists in between the subsample sets and the original sample set, this paper proposes a randomoriented partition method based on distributed k-means clustering. Not that the method is used to take a layer of the training results directly as input of the next layer, but that the k-means clustering algorithm is used to cluster into the number of the next layer node clusters. Then, the paper divides each cluster samples into N parts randomly, and takes out one from each cluster reassembled into N subsample sets to next layer of training to ensure the distribution of the subsample sets similar to original sample set. The results show that this method can not only improve learning ability effectively, but also reduce the jitter of training model.