An efficient parallel minimum spanning tree is proposed based on the classical Borüvka's algorithm on message passing parallel machine. Three methods were used to improve its efficiency, including two-phase union and packaged contraction for reducing communication costs, and the balanced data distribution for computation balance in each processor. The computation and communication costs of the algorithm are O(n²/p) and O((t_sp+t_wn)n/p). On Dawning-1000 parallel machine, it gets a speedup of 12 on 16 processors with a sparse graph of 10 000 vertices.