This paper studies the critical traffic matrices selection problem and develops an algorithm called MinMat which uses information entropy to select the first critical matrices at first, then takes merging cost into consideration when agglomerating a pair of clusters. The algorithm is evaluated by using a large collection of real traffic matrices collected in Abilene network. Theoretical analysis and experimental results demonstrate that MinMat algorithm is more effective than K-means, Hierarchical Agglomeration, CritAC, and by simulating on Totem, it is concluded that a small number of critical traffic matrices suffice to yield satisfactory performance.