In this paper, we study the joint optimization problem of computation path selection and workload allocation for in-network computing at the edge. The existing works, which only consider the end-to-end latency, ignore the operational cost of the servers and the dynamic tasks arrived in an online manner. Thus, we investigate the first online scheduling problem of path selection and workload allocation for in-network computing. Such a problem is modeled as a mixed integer programming problem, which tries to jointly minimize the server operating cost and end-to-end latency. Then, a dynamic 3D-map is constructed to take both the operation cost and end-to-end latency into account. Based on the constructed 3D-map, a cooperative learning based algorithm with ant colony optimization is proposed. Finally, the extensive simulation demonstrates the proposed algorithm shows good robustness and outperforms the state-of-the-art algorithms.