Optimal power flow is usually a non-convex problem of central coordination because of the nonlinear relations between nodal voltages and power supplied/withdrawn. In this paper, we propose a sequential distributed algorithm to address the problem of cost minimization power flow in distribution systems. Each node is considered as an agent, which solves its local optimization problem with the local knowledge of the network and communicated state variables of its neighbors. The power generation is optimized by the decomposed sub-problem at each node in backward sequence and the voltages are calculated in forward sequence. Our approach does not require any form of central coordination or regional control. The proposed algorithm has a fast convergence which is illustrated on various distribution test systems.