The advanced regional E-grid η-coordinates model(AREM) is used to simulate the persistent heavy rain in South China during the period of June 20 to 22, 2005, which caused a historical flood, and the AREM-simulated precipitation sensitivities to initial and boundary fields are tested. Results show that there are obvious differences between initial and boundary conditions' impact on simulated precipitation in the different time periods of simulation. The larger the initial disturbance error is, the larger is the model error. The experiments also reveal that forecast error firstly rapidly grows in the meso-and small scale systems accompanied with unstable moist convection process, then propagates to large scale ones. Because of the gradual release of convective available potential energy, the atmospheric instability reduces, and the error grows slowly in the large scale system. However, in the case that the initial and boundary fields are of the same precision, it is better to add the boundary mesoscale information, especially below 400 km, rather than to add the initial mesoscale information for effectively suppressing the growth of errors.