In recent years, fast processing for extremely large-scale graph, consisting of millions to trillions of vertices and billions to hundreds of trillions of edges, is becoming increasingly important. GIM-V graph processing algorithm based on MapReduce, which automatically manages petabyte-scale data, is designed as general graph processing method. Besides, recent large-scale computing systems tend to employ GPUs to gain good peak performance and high memory bandwidth. However, acceleration factor using GPU for large graph processing and the way of efficient data distribution are not clear. We implemented multi GPU based GIM-V system and investigated the effect of graph partitioning, which is a method to reduce inter-node data transfer. Our experiment showed that GPU performed 7.17x faster than CPU on Map but not on Shuffle and Reduce, and graph partitioning reduced data transfer time by 54% but total elapsed time increased due to workload imbalance.