In the big data era, the video data of social media increase rapidly. To detect and block pornographic videos, traditional pornographic image detection methods cannot be applied directly to large-scaled video data. For this purpose, a parallel computing network has been set up by a lot of cheap computers for massive pornographic video data detection. First, we propose a key-frame extraction algorithm based on inter-frame similarity. This algorithm only uses the local information of the video and can be dispatched to multiple computers for parallel processing. The results of key-frame extraction are persisted to the distributed file system. Next, in order to determine whether a video contains pornographic key-frames, we propose a discriminative multiple Gaussian mixture models to extract skin color regions and an active relevance feedback bootstrap algorithm to detect the face. Finally, the geometric characteristics of the body are used to determine whether the key-frame is a pornographic image, and according to the number of pornographic key-frame in the video to decide whether the video is pornographic or not. Compared with some existing methods, the detection accuracy has been greatly improved. Because of the proposed methods are processed in different computer nodes for parallel computing, the processing speed is only related to the scale of the video data and the number of the computers. In practical applications, it can meet demands only need to select enough computers according to the scale of the video data. In theory, it can be used for video data at any scale.