In this paper, we propose a new metric index, called M+ -tree, which is a tree dynamically organized for large datasets in metric spaces. The proposed M+ -tree takes full advantages of M-tree and MVP-tree, with a new concept called key dimension, which effectively reduces response time for similarity search. The main idea behind the key dimension is to make the fanout of tree larger by partitioning a subspace further into two subspaces, called twin-nodes. We can double the filtering effectiveness by utilizing the twin-nodes. In addition, for the purpose of ensuring high space utilization, we also conduct data reallocation between the twin nodes dynamically. Our experiment shows that higher filtering efficiency can be obtained by using the key dimensions for r-neighbor search and k-NN (k-nearest neighbor). We will report our experimental results in this paper.