Identifying densely connected components in a graph (network) is known as the graph clustering or community detection problem. Modularity is a metric that measures the quality of partitioning, and algorithms that optimize modularity have been demonstrated to not only generate high quality solutions but also to converge quickly. Louvain is popular serial algorithm based on the concept of modularity optimization. We developed different heuristics to parallelize the Louvain algorithm and developed implementations targeting multi-GPU systems. The key features of our multi-GPU implementation are: a modified data structure to enable efficient computations on GPU systems, efficient sharing of work between CPUs and GPUs, and the use of cooperative groups for efficient execution on the GPUs.