Large-scale graph processing is becoming more important due to the large vol-ume of available data, the low cost of storage and the stunning success of onlinesocial networks. Besides, recent large-scale computing systems tend to employGPUs to gain good peak performance and high memory bandwidth. Mars isone of the MapReduce library accelerated with GPUs;however, the problemson GPGPU-based large-scale graph processing, such as the performance im-provement by using GPUs, the memory management in GPUs, and the datadistribution between multiple GPUs, are not investigated. In order to clarifythe problems, we implemented several graph processing applications such asPageRank, Random Walk with Restart(RWR) and Connected Components ontop of Mars and compared the results with PEGASUS, one of the large-scalegraph processing library. Our experimental results show that the mean time ofone iteration in Mars is 2.17-9.53 times faster on PageRank, 2.18-5.47 times onRWR, and 2.51-8.46 times on Connected Components respectively comparedwith PEGASUS.