To address the problem of load balance in the distributed crawler system, we propose a new load balancing algorithm. First, we investigate the impact of different features among nodes of distributed crawler system on running time. Based on these features, a runtime model is established. Second, the minimum variance of running time of each node predicted by the model is regarded as the load-balancing objective function. Finally, we utilize an approximate gradient descent (AGD) to optimize the objective function and produce a sequence of task allocation to the nodes. The experimental results show an improvement in the running time of the distributed crawler system.