Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

GraphSnapShot: Caching Local Structure for Fast Graph Learning

Dong Liu, Roger Waleffe, Meng Jiang, Shivaram Venkataraman

TL;DR

GraphSnapShot tackles the challenge of training on large, dynamic graphs by maintaining a centralized cache of local, $k$-hop subgraphs and a hybrid static-dynamic sampling strategy. The framework combines static preprocessing with dynamic updates, using a multi-level cache hierarchy and a mix of FCR, OTF, and shared caching to minimize disk I/O while preserving topology accuracy. Empirical results on ogbn benchmarks show substantial training time reductions and GPU memory savings compared to baselines, underlining the method's scalability and practicality for dynamic graph learning. Overall, GraphSnapShot offers a disk-cache-memory approach that intelligently balances sampling quality and computational cost, enabling efficient learning on evolving graphs in real-world applications.

Abstract

In our recent research, we have developed a framework called GraphSnapShot, which has been proven an useful tool for graph learning acceleration. GraphSnapShot is a framework for fast cache, storage, retrieval and computation for graph learning. It can quickly store and update the local topology of graph structure and allows us to track patterns in the structure of graph networks, just like take snapshots of the graphs. In experiments, GraphSnapShot shows efficiency, it can achieve up to 30% training acceleration and 73% memory reduction for lossless graph ML training compared to current baselines such as dgl.This technique is particular useful for large dynamic graph learning tasks such as social media analysis and recommendation systems to process complex relationships between entities. The code for GraphSnapShot is publicly available at https://github.com/NoakLiu/GraphSnapShot.

GraphSnapShot: Caching Local Structure for Fast Graph Learning

TL;DR

GraphSnapShot tackles the challenge of training on large, dynamic graphs by maintaining a centralized cache of local, -hop subgraphs and a hybrid static-dynamic sampling strategy. The framework combines static preprocessing with dynamic updates, using a multi-level cache hierarchy and a mix of FCR, OTF, and shared caching to minimize disk I/O while preserving topology accuracy. Empirical results on ogbn benchmarks show substantial training time reductions and GPU memory savings compared to baselines, underlining the method's scalability and practicality for dynamic graph learning. Overall, GraphSnapShot offers a disk-cache-memory approach that intelligently balances sampling quality and computational cost, enabling efficient learning on evolving graphs in real-world applications.

Abstract

In our recent research, we have developed a framework called GraphSnapShot, which has been proven an useful tool for graph learning acceleration. GraphSnapShot is a framework for fast cache, storage, retrieval and computation for graph learning. It can quickly store and update the local topology of graph structure and allows us to track patterns in the structure of graph networks, just like take snapshots of the graphs. In experiments, GraphSnapShot shows efficiency, it can achieve up to 30% training acceleration and 73% memory reduction for lossless graph ML training compared to current baselines such as dgl.This technique is particular useful for large dynamic graph learning tasks such as social media analysis and recommendation systems to process complex relationships between entities. The code for GraphSnapShot is publicly available at https://github.com/NoakLiu/GraphSnapShot.

Paper Structure

This paper contains 48 sections, 29 equations, 9 figures, 10 tables, 3 algorithms.

Table of Contents

  1. Introduction
  2. Background and Motivation
  3. Local Structure Caching in Graph Learning
  4. Challenges in Dynamic Graph Learning
  5. Model Construction
  6. Centralized Cache for Local Structure
  7. Static and Dynamic Sampling
  8. Cache Management Strategies
  9. GraphSnapShot Architecture
  10. System Components
  11. Hybrid Caching Mechanism
  12. Static Sampling.
  13. Dynamic Resampling.
  14. Hierarchical Cache Design
  15. Dynamic and Static Sampling Trade-off
  16. ...and 33 more sections

Figures (9)

  • Figure 1: GraphSnapShot Model Architecture Code: https://github.com/NoakLiu/GraphSnapShot
  • Figure 2: System Architecture of GraphSnapShot. Disk stores the complete graph structure, cache accelerates graph learning, and CPU/GPU performs computations.
  • Figure 3: Performance Comparison on ogbn-arxiv
  • Figure 4: Performance Comparison on ogbn-products
  • Figure 5: Performance Comparison on ogbn-mag
  • ...and 4 more figures