Yunmo Zhang and Jiacheng Huang, City University of Hong Kong; Xizhe Yin, Independent Researcher; Junqiao Qiu, City University of Hong Kong; Hong Xu, The Chinese University of Hong Kong; Chun Jason Xue, Mohamed bin Zayed University of Artificial Intelligence
Large-scale evolving graph analytics (EGA), which evaluates graph queries over sequences of snapshots, is facing growing demands for real-time insight extraction. While GPUs offer immense potential for accelerating graph workloads, they suffer from the memory capacity wall and poor hardware utilization when applied to EGA.
To bridge this gap, this work presents POEGA, a GPU-centric framework for efficient large-scale EGA. The core idea of POEGA is to leverage proxy graphs to minimize out-of-memory IO. It achieves this by first analyzing a compact in-memory graph abstraction to drive approximate results, thereby guiding the out-of-memory refinement. Although this approach incurs more computations, our key insight is that this cost can be amortized by exploiting the GPU’s massive parallelism to process multiple snapshots concurrently. This concurrency is supported by a carefully designed fused kernel and a novel bound-based pruning technique. Furthermore, we address a commonly overlooked memory bottleneck caused by multi-version vertex states, which arises when scaling concurrent analysis to a large number of snapshots, by introducing an adaptive state compaction format. Evaluation shows that POEGA yields 3.7-23.5× speedups over the state-of-the-art EGA solutions across a range of real-world datasets.
