Real world large scale networks exhibit intrinsiccommunity structure, with dense intra-community connectivityand sparse inter-community connectivity. Leveraging their communitystructure for parallelization of computational tasks andapplications, is a significant step towards computational efficiencyand application effectiveness. We propose a weighted depth-firstsearchgraph partitioning algorithm for community formationthat preserves the needed community dependency without anycycles. To comply with heterogeneity in community structure andsize of the real world networks, we use a flexible limiting valuefor them. Further, our algorithm is a diversion from the existingmodularity based algorithms. We evaluate our algorithm as thequality of the generated partitions, measured in terms of numberof graph cuts.