I/O performance for very large-scale storage systems has become more and more important as the need for storage has grown dramatically. Replication is one of the most important approaches to improve the I/O performance of massive distributed storage system. The storage node to which the Read/Write request is directed is the critical issue when multiple replicas exist in some different storage nodes simultaneously. This paper proposes and evaluates a novel master OSD model, called Dynamic Selectable Master OSD (DSMOSD), which allows client to direct the Read/Write request to the optimized OSD. This approach has the salient advantage of simultaneously improving I/O performance and balancing the loads of all the OSDs. This paper first describes the architecture of Master OSD and characterizes the advantages of this model contrasting to the pseudo random hash function. Then an improved model, which is considered as DS-MOSD, is presented and evaluated. The experimental results show that the load unbalance degree and network overheads of DS-MOSD model are smaller than that of Fixed MOSD model, while the success probability of Write/Update transactions of DS-MOSD model is larger than that of Fixed MOSD model.