RealGraphGPU++: A High-Performance GPU-Based Graph Engine with Direct Storage-to-DM IO

초록

Recently, with the increasing size of real-world networks, graph engines have been studied extensively for efficient graph analysis. As one of the state-of-the-art single-machine-based graph engines, RealGraphGPU processes large-scale graphs very efficiently thanks to its well-designed architecture and the strong parallel-computing power of GPU. Via a preliminary analysis, we first observe RealGraphGPU has a good chance for more performance improvement in IOs between storage and GPU’s device memory. This motivates us to present RealGraphGPU++, a solution that substantially reduces IO time by establishing a direct data path between storage and device memory. Additionally, it employs asynchronous processing of CPU and GPU tasks to issue IO requests more frequently, thereby improving overall performance by achieving higher IO bandwidth. Experimental results on real-world datasets show that RealGraphGPU++ outperforms dramatically existing 11 state-of-the-art graph engines including RealGraphGPU

키워드