Abstract: Taking performance similarity as a point of departure, this paper proposes an algorithm for buffer library compression. Concepts of virtual cell, real cell and probability weighted distance are established. Buffers with best performance under certain environment points are identified as cells of virtual library. Virtual cells are clustered in the second step. Center of each cluster is finally mapped to compressed library. Environmental adaptability of each buffer is quantized to give priority to cluster center identification. Above idea is achieved by using three programs. Result shows that three implementations achieve 9．6%, 10.4% and 6% of SSE respectively compared with existed algorithm, which represents better performance similarities between compressed library and the unpruned one.