Surface roughness is a universal characteristic of natural faults. Roughness can be represented statistically as a random field that is approximately self-similar over many orders of magnitude in scale-length with a ratio of amplitude to length scale that typically falls into the range 10-3 to 10-2. Incorporating realistic rough fault surface into 3D numerical simulations of earthquake dynamic rupture provides guidance to build kinematic rupture generator. We have built a database of more than 1000 simulations of 3D strike slip dynamic rupture for different realizations of rough fault surfaces at different fault roughness levels. We first have explored the role of the fault roughness in influencing the 1-point and 2-point statistics of earthquake source parameters such as rupture velocity, peak slip rate, total slip, and slip rise time. Fault roughness reduces the amplitudes of rupture velocity, peak slip rate, rise time and total slip. Then, we have extended our study to assess supershear transition mechanisms that operate in 3D on rough faults and what factors contribute to the frequency of occurrence and spatial extent of supershear rupture episodes. We have reconciled the conflict that the supershear is favored by fault roughness from 2D numerical simulation (Bruhat et al., 2016) and unfavored by fault roughness from field observations (Bouchon et al., 2010) by dividing supershear into two types supershear transitions: free surface supershear transition and buried supershear transition. Finally, we have investigated how well simulated earthquake behaviors on rough fault relate to direct geological observations, such as free surface lateral slip, plastic strain, and shallow slip deficits. We have found that fault roughness in the form of a power law leads to self-affine surface lateral slip, which is in agreement with recent optical imaging observations in the 1992 Landers earthquake (Milliner et al., 2015). Also fault roughness tends to produce individual events with large shallow slip deficits, which may help explain the suggestion that has been made, based on a small number of earthquakes, that the SSD tends to be larger on immature faults.