How to solve the Non-IID problem in Federated learning if we do not care about the communication overhead?

It seems that recent studies focued on how to minimize the communication overhead of Federated Learning and also improve the accuracy of models when facing NIID problems. However, compared with MPC, FL is quite effective. Thus, I am wondering whether there exist a solution for Non-IID problem in Federated learning if we do not care about the communication overhead?