Abstract: As a computation model, LogP provides us the algorithm design criteria independent of the concrete system. Although LogP model can schedule the communication and computation precisely, algorithm design is involved in complicated details. The practical efficiency gained under the LogP model may be low. We improve LogP model in both algorithm design and model characterization by introducing barrier synchronization and long message, which not only exploits the performance of MPP, but also makes the prediction be close to the runtime effect. We analyze and verify our methodology in an FFT parallel algorithm.