As the scale of network-on-chip (NoC) expands and related research goes further, task scheduling becomes a key problem for optimizing system temperature. For the task scheduling problem of NoC, we present a task scheduling method based on the shortest Manhattan distance. The method fully considers the shortest Manhattan path between the pairs of communication nodes in a core communication graph, uses a search algorithm to find the destination node for scheduling tasks, and determines the task scheduling pair with simulated annealing algorithm. SMDS experimental results indicate that, compared with the traditional distributed task migration (DTM) strategy, for 6*6, 8*8 and 10*10 topologies, the optimization for the times of migration is reduced by 22.08%, 21.74% and 23.02%, respectively, and the average optimization rate for average hop count is 24.04%, 29.18% and 23.04%, which achieves system temperature optimization.