A novel supervised linear dimensionality reduction algorithm called margin discriminant projection (MDP) is proposed to extract low-dimensional features with good performance of discriminant. MDP aims to minimize maximum distance of samples belong to the same class and maximize minimum distance of samples belong to different classes, and at the sametime preserve the geometrical structure of data manifold. Compared with classical algorithms based on the definition of margin, MDP is good at preserving the global properties, such as geometrical and discriminant structure of data manifold, and can overcome small size sample problem. Due to its low cost of computation, MDP can be directly applied on ultra-high dimensional big data dimensionality reduction. Experimental results on five face data sets show its effectiveness for feature extraction on big data.