To facilitate the communication between remote and adjacent nodes on the large-scale Networks-on-Chip (NoC), a hierarchical Cluster-like Hierarchical Mesh (CHM) topology based on region partition was proposed. Correspondingly, to avoid performance degradation due to the bad network congestion near the intermediate nodes, an adaptive source region path selection algorithm was elaborated. According to the region characteristic in CHM, the routing decision was determined in the source region instead of source node only, and the adaptive routing node pair was distinguished among bottom and upper node pairs, which enhanced the routing selection performance for those node pairs, so as to alleviate the bad network congestion. The experimental results show that, compared with the shortest path algorithm, the proposed algorithm can increase saturation injection by at most 51% and 31% respectively under synthetic and local traffic patterns, showing an effective improvement in network throughput.