MSPipe: Minimal Staleness Pipeline for Efficient Temporal GNN Training
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: Representation Learning, Temporal Graph Neural Networks, Asynchronous Training, Staled Memory
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: Temporal graph neural networks (TGNNs) have demonstrated exceptional performance in modeling interactions on dynamic graphs. However, the adoption of memory modules in state-of-the-art TGNNs introduces significant overhead, leading to performance bottlenecks during training. This paper presents MSPipe, a minimal staleness pipeline design for maximizing training throughput of memory-based TGNNs, tailored to maintain model accuracy and reduce resource contention. Our design addresses the unique challenges associated with fetching and updating memory modules in TGNNs. We propose an online pipeline scheduling algorithm that strategically breaks temporal dependencies between iterations with minimal staleness and delays memory fetching (for obtaining fresher memory vectors) without stalling the GNN training stage or causing resource contention. We further design a staleness mitigation mechanism to improve training convergence and model accuracy. We provide convergence analysis and demonstrate that MSPipe retains the same convergence rate as vanilla sampling-based GNN training. Our experiments show that MSPipe achieves up to 2.45$\times$ speed-up without sacrificing accuracy, making it a promising solution for efficient TGNN training. The implementation (anonymous) for our paper can be found at [https://anonymous.4open.science/r/MSPipe/](https://anonymous.4open.science/r/MSPipe/).
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 4465
Loading