CPU-GPU cluster computing systems are widely used to process large-scale data in various fields owing to its famous thousands of computing cores and high computation intensiveness. However, its high performance is always accompanied by high power consumption, even further resulting in reduced reliability and instantaneous failure. In this paper, a reliable cluster Resource Scheduling approach with dataflow Natural Attribute Priority (RSNAP) is proposed, and it is designed to increase the reliability of collaborative computing nodes. On the basis of the idea of genetic algorithm, it searches for an optimal task-processor allocation scheme with considering the characteristics of tasks and processors (GPUs). For tasks that have data dependencies, we reduce the load that needs to be performed by adjusting the frequency and voltage of the GPU to further reduce node power consumption without affecting the natural distribution of data in the nodes. With less loads, GPU power consumption pressure has been eased, further enhancing system reliability. While for tasks that have no dependencies, we improve system reliability through dynamic task migration. The experiment results show that RSNAP can reduce GPU power consumption and improve reliability of collaborative computing.