Autonomous, independent management of dynamic graphs on GPUs

In this paper, we present a new, dynamic graph data structure, built to deliver high update rates while keeping a low memory footprint using autonomous memory management directly on the GPU. By transferring the memory management to the GPU, efficient updating of the graph structure and fast initialization times are enabled as no additional memory allocation calls or reallocation procedures are necessary since they are handled directly on the device. In comparison to previous work, this optimized approach allows for significantly lower initialization times (up to 300× faster) and much higher update rates for significant changes to the graph structure and equal rates for small changes. The framework provides different update implementations tailored specifically to different graph properties, enabling over 100 million of updates per second and keeping tens of millions of vertices and hundreds of millions of edges in memory without transferring data back and forth between device and host.