ИСТИНА

Systems equipped with modern NVIDIA GPUs nowadays are capable of solving many real-world problems, including large-scale graph processing. Efficiently implementing graph algorithms on GPUs is a challenging task, since modern real-world graphs have irregular structure. Current paper describes approaches, which can be used for developing efficient implementation of label propagation algorithm, frequently used to solve graph clustering and community detection problems. Compared to already existing GPU-based implementations, new optimization techniques have been proposed, including graph preprocessing, efficient load-balancing, using unified memory for out-of-core graph processing, and several others. The performance of the developed implementationhas been evaluated on synthetic and medium-scaled real-world graphs, resulting in up to 2 times better results compared to existing approaches.