In scenarios where client data are non-independent and identically distributed (Non-IID), this paper proposes a communication-efficient personalized federated learning algorithm based on Non-IID data to provide personalized and efficient communication solutions for clients. Specifically, to leverage the knowledge among similar clients and retain personalized information of local clients, we develop a personalized federated learning algorithm that combines hierarchical modeling with clustering ideas. Furthermore, to address the issue of high communication overhead, we design a selective model aggregation strategy. The central server evaluates the similarity between the client data distribution and the global data distribution using maximum mean discrepancy. Based on this similarity, the server computes a priority score for each client and selects those with higher scores for communication. This strategy effectively reduces the cumulative communication rounds between clients and the central server, thereby improving communication efficiency and accelerating model convergence. Experimental results demonstrate that compared with existing representative works, the proposed algorithm reduces the cumulative commu nication rounds by over 50% while maintaining high accuracy.