CNN-based methods have achieved notable performance in the research of single image super resolution domain. To further improve the representation ability and performance of networks, most research works have adopted the attention mechanism. In this survey, we introduce a taxonomy for the attention based super-resolution networks and classify existing methods into two categories: first-order and secondorder attention. We also provide comparisons between the models in terms of network scale, memory footprint, type of network losses and important architectural differences for attention implementation. An analysis tool from recent network interpretation works is applied to verify the improvements of the evolving attention mechanism. Finally, we analyze and discuss challenges in processing real degraded images, and point out the problems and potential topics in future research work.