Towards Sufficient GPU-accelerated Dynamic Graph Management: Survey and Experiment

Dynamic graph management (DGM) systems are designed to effectively handle changing graph data, which is a fundamental problem for many graph-based applications. Recently, researchers have designed GPU-based solutions for DGM and its downstream applications, thanks to GPUs' massive parallelism power. However, there is a lack of universal models that summarize the features and design principles of GPU-accelerated DGM systems. Additionally, existing studies test GPU-based DGM systems without unified metrics and workloads. Under this circumstance, we propose a conceptual model for GPU-accelerated DGM to demonstrate a DGM system's components, key primitives, and optimization choices. Next, we evaluate six representative systems, testing their update and query performance with unified metrics and workloads of different algorithmic behaviors. We also extend existing systems to seek insight to fill the current research gap in multi-GPU support, concurrency control, resource utilization, and so on. Our evaluation yielded new insights on the pros and cons of different systems: (1) Hashing-based systems perform best for graph updates but may not be suitable for all applications. (2) Finding a system that fits all workloads is challenging, and hybrid data storage may be a solution. (3) To select the most suitable DGM system for a specific workload, it is essential to consider hardware-related metrics. Finally, we provide recommendations and suggestions for future studies based on our experimental results and observations.