On Optimizing the Communication of Model Parallelism