Abstract: We present a novel load-balancing algorithm to the distributed system. In the algorithm, tasks are classified by the resource which they would take up later. While scheduling, every node's load is calculated dynamically through the task type and the resource's overhead collected periodically. Then a suitable node is found to the task in a subset of the whole nodes. The algorithm can remove the herd effect caused by delayed information, and can decrease the mean response time of a few tasks that have special requirement in this regard. It has been adopted in a distributed operation system developed by our group.