DeltaGNN: Accelerating Graph Neural Networks on Dynamic Graphs With Delta Updating

Graph neural network (GNN) accelerators haveachieved prominent performance speedup on static graphs butfallen with inefficiency on dynamic graphs. The reason is that indynamic graphs, updating on a few vertices will introduce enor-mous redundant neighbor reaggregation and feature reupdating.Moreover, evolving graph structure makes graph preprocessingimpractical and incurs random memory accesses which can onlybe determined at runtime. In this article, we propose DeltaGNN,an algorithm and accelerator co-design for GNN acceleration ondynamic graphs. In algorithm, we first propose a delta updatingalgorithm, which identifies the sensitivity of vertices and reducesthe aggregation and updating operations of insensitive verticeswithout accuracy compromise. In hardware, we propose a novelsensitivity remapping cache to satisfy the dissimilar reusabilityof vertices under different sensitivity without preprocessingrequirement. To tackle the workload imbalance, we implementfeature-disperse execution to support different feature updatingbetween sensitive and insensitive vertices. Moreover, we introducevertex feature coalescing to reduce the amount of feature vectorsby exploiting the locality within vertex accesses. We then proposelightweight yet effective hardware optimizations to make ourdesign applicable to conventional GNN accelerators. Comparedto the state-of-the-art GNN accelerators, our DeltaGNN gains anaverage of 1.5x-11.8xspeedup and 1.3x-8.6xenergy efficiencyimprovement on dynamic graphs