A sequential minimal optimization(SMO)algorithm is a very effective method which can improve the training speed of support vector machine(SVM). However,the SMO algorithm is still quite slow in the large?scale datasets. In order to increase the training speed,an optimization strategy which can maintain the training accuracy is proposed. The strategy is to skip the part of the vector irrelevant to accuracy,prematurely finish cycle and relaxe KKT conditions so that it can shrink the working set. The results show that this strategy can significantly reduce the training time and the accuracy is still high in several datasets.