.png){kind=icon}
.png){kind=icon}
Return
Torch-feat: GNN sampling training data loader based on feature data extraction operators
.png){kind=icon}
PRE
AI
DOI：10.1007/s42514-025-00267-6.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
When training Graph Neural Networks (GNNs) on large-scale graph datasets, GNN systems such as PaGraph utilize caching strategies to accelerate data loading. However, this strategy encounters significant loading overhead when cached node data is not found. To address this issue, we propose Torch-feat, a GNN sampling training data loader based on a feature data extraction operator. First, PaGraph's self-contained graph partitioning algorithm is enhanced, ensuring that each training process handles data in its own partition, thereby avoiding cross-process data transmission. Second, two-level storage (CPU lockstep memory and GPU memory) is employed to store feature data, with feature data placed in GPU memory to prevent redundant loading. Finally, we propose a feature data extraction operator, where the GPU concurrently extracts and rearranges node feature data from both GPU and CPU lockstep memory, eliminating the need for CPU involvement in data extraction and minimizing CPU resource contention due to concurrent feature data extraction tasks. Comparative experiments with DGL, PaGraph, and DGL (prefetch) in a CPU+GPU environment demonstrate that Torch-feat accelerates data loading by 5.98x, 4.22x and 1.48x respectively, when training a GNN model on the ogbn-papers100M dataset.
Keywords:
Graph neural networks
Large graphs
Data loading
Mini-batch training
Extraction operators
Journal
C
IF:
1.9
Papers:
38
Citations:
253