The paper studies multi-period production control in a centralized fully flexible manufacturing system where parts are dynamically dispatched from a central control station to n local work stations, each with random processing rate. After being processed at local stations, parts must return to the central station where a variable portion of the returning parts are either discarded as scrap or routed back to the system for rework. A dynamic optimal dispatching model is employed to determine the short-term production control policy which minimizes total expected operational costs over a finite horizon. For a non-feedback case, the paper shows that the dynamic recursive functions are separable across the stations. Thereby, a decentralized optimal control scheme can be established. For a feedback case, the station separability does not necessarily hold. A solution procedure for the feedback case is developed based on convex analysis.