Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Incremental GNN Embedding Computation on Streaming Graphs

Qiange Wang, Haoran Lv, Yanfeng Zhang, Weng-Fai Wong, Bingsheng He

Abstract

Graph Neural Network (GNN) on streaming graphs has gained increasing popularity. However, its practical deployment remains challenging, as the inference process relies on Runtime Embedding Computation (RTEC) to capture recent graph changes. This process incurs heavyweight multi-hop graph traversal overhead, which significantly undermines computation efficiency. We observe that the intermediate results for large portions of the graph remain unchanged during graph evolution, and thus redundant computations can be effectively eliminated through carefully designed incremental methods. In this work, we propose an efficient framework for incrementalizing RTEC on streaming graphs.The key idea is to decouple GNN computation into a set of generalized, fine-grained operators and safely reorder them, transforming the expensive full-neighbor GNN computation into a more efficient form over the affected subgraph. With this design, our framework preserves the semantics and accuracy of the original full-neighbor computation while supporting a wide range of GNN models with complex message-passing patterns. To further scale to graphs with massive historical results, we develop a GPU-CPU co-processing system that offloads embeddings to CPU memory with communication-optimized scheduling. Experiments across diverse graph sizes and GNN models show that our method reduces computation by 64%-99% and achieves 1.7x-145.8x speedups over existing solutions.

Incremental GNN Embedding Computation on Streaming Graphs

Abstract

Graph Neural Network (GNN) on streaming graphs has gained increasing popularity. However, its practical deployment remains challenging, as the inference process relies on Runtime Embedding Computation (RTEC) to capture recent graph changes. This process incurs heavyweight multi-hop graph traversal overhead, which significantly undermines computation efficiency. We observe that the intermediate results for large portions of the graph remain unchanged during graph evolution, and thus redundant computations can be effectively eliminated through carefully designed incremental methods. In this work, we propose an efficient framework for incrementalizing RTEC on streaming graphs.The key idea is to decouple GNN computation into a set of generalized, fine-grained operators and safely reorder them, transforming the expensive full-neighbor GNN computation into a more efficient form over the affected subgraph. With this design, our framework preserves the semantics and accuracy of the original full-neighbor computation while supporting a wide range of GNN models with complex message-passing patterns. To further scale to graphs with massive historical results, we develop a GPU-CPU co-processing system that offloads embeddings to CPU memory with communication-optimized scheduling. Experiments across diverse graph sizes and GNN models show that our method reduces computation by 64%-99% and achieves 1.7x-145.8x speedups over existing solutions.
Paper Structure (25 sections, 1 theorem, 9 equations, 12 figures, 7 tables, 4 algorithms)

This paper contains 25 sections, 1 theorem, 9 equations, 12 figures, 7 tables, 4 algorithms.

Table of Contents

  1. Introduction
  2. Background
  3. GNN Basics
  4. GNN Training and Inference for Streaming Graphs
  5. $$ RTEC for Streaming Graphs
  6. The Redundant Computation of $$ RTEC
  7. Limitation of Non-incremental Solutions
  8. Incremental $$ RTEC: Opportunity and Challenges
  9. Incremental $$ RTEC Framework
  10. Fine-grained Operator Decoupling for Incremental $$ RTEC
  11. Reordered Incremental $$ RTEC Workflow
  12. Equivalence Analysis
  13. Application to Various GNN Models
  14. NeutronRT System
  15. Incremental Computation Engine
  16. ...and 10 more sections

Key Result

Theorem 1

Equivalence of full-neighbor $$RTEC and incremental $$RTEC Given a vertex $v$, its original neighborhood $N$, an update set $\Delta N$, and the corresponding vertex embeddings $\{ \textbf{h}^l_u \mid u \in N \cup \Delta N \}$, the output embedding computed by Algorithm alg:inc_comp, using the increm

Figures (12)

  • Figure 1: $$RTEC on a toy graph with edge updates $\langle4,2\rangle$ and $\langle4,4\rangle$ using a 2-layer model. Blue items in (c) indicate redundant computations on the unaffected subgraph.
  • Figure 2: Processed edge volume is normalized to the Affected Subgraph (AS). FN, NS#, and UER denote $$RTEC with full-neighbor computation, neighbor sampling, and unaffected embedding reuse, respectively. The percentage of redundant computation on unaffected subgraphs is labeled above each bar, except for NS approaches, which compute on both sampled affected and unaffected subgraphs.
  • Figure 3: An illustration of $$RTEC-NS and $$RTEC-UER and $$RTEC-Inc for the affected vertex v2. $$RTEC-NS and $$RTEC-UER cannot eliminate redundant computation on unaffected subgraphs.
  • Figure 4: Full-neighbor $$RTEC and incremental $$RTEC for the GIN model. Message: $f(x)=x$; Aggregate:sum(); update: MLP().
  • Figure 5: A graphical illustration of full-neighbor $$RTEC and the reordered incremental $$RTEC. Colored boxes represent result tensors and their compositions from upstream operations; solid arrows indicate data flow, while dashed arrows denote the reuse paths.
  • ...and 7 more figures

Theorems & Definitions (1)

  • Theorem 1