With the ever-increasing scale and complexity of social network and online business, Recommender Systems (RS) have played crucial roles in information processing and filtering in various online applications, although suffering from such as data sparsity and low accuracy problems. Meanwhile, recent researches try to enhance the performance of RS through such social network and clustering algorithms, however, they may fail to achieve further improvement in large-scale online recommendation due to the serious information overload. In this article, a novel social recommendation approach with online update referred to as LsRec is proposed, which generally contains offline computation and online incremental update. More precisely, LsRec not only takes account of user's social relationship, but also clusters items according to the similarity degree, furthermore, LsRec performs recommendation in each generated cluster respectively. In practice, LsRec could be capable of exploiting user-level social influence, and capturing the intricate relationship between items. In addition, theoretical proof could provide convergence guarantee for the model. Specifically, with the appealing merit of flexible online update scenario, LsRec could yield high performance in large-scale online recommendation with low computational complexity. Extensive experimental analysis over four real world datasets demonstrate the effectiveness and efficiency of LsRec, which indicates that LsRec could significantly outperform state-of-the-art recommender approaches, especially in large-scale online recommendation.