Skip to Main Content
 

Global Search Box

 
 
 

ETD Abstract Container

Abstract Header

Structured Pruning Graph Neural Networks Models for Inference Acceleration

Naghibi Saleh, Gelareh
http://orcid.org/0009-0006-9483-412X
http://rave.ohiolink.edu/etdc/view?acc_num=ysu173375500625166

Abstract Details

2024, Master of Computing and Information Systems, Youngstown State University, Department of Computer Science and Information Systems.
As graphs used as input for real-world applications become larger, training and inference workloads for Graph Neural Networks (GNNs) can quickly become infeasible under any reasonable latency requirements. Unlike traditional Deep Neural Networks (DNNs), GNNs confront various computational problems due to their large dimensions, unstructured nature, and sparsity in the input graphs. Several pruning mechanisms have been used in the past, including both irregular and structured methods, to reduce the size of GNN models while maintaining performance. The challenge is that some pruning algorithms, particularly those that are irregular or unstructured, do not help improve the computing performance on parallel architectures because the sparse and irregular connections impose limitations on the amount of parallelism in the dense matrix multiplication kernels. The structured pruning methods reduce the size of GNN models however, their adoption has been somewhat limited due to the inability of current structured model pruning techniques to leverage GPU resources to their fullest, especially with vectorized hardware. This work shows that the newly introduced structured pruning method enhances the scalability of GNNs on modern parallel architectures. The method prunes redundant connections in a structured way, optimizing GNNs for modern GPUs and improving parallelism for efficient GPU utilization. The results show that the structured pruning method achieves high sparsity and reduced computing time without sacrificing accuracy. In real-time applications where latency is critical, such as social networks and autonomous driving, these results allow for the effective deployment of GNNs.
Alina Lazar, PhD (Advisor)
Feng Yu, PhD (Committee Member)
Robert Gilliland, PhD (Committee Member)
47 p.
Computer Science; Information Systems
Graph Neural Networks; GNN; Structured Pruning; Inference Acceleration; Real-Time-Large-Scale Graph; GCN; GIN; GAT

Recommended Citations

Citations

  • Naghibi Saleh, G. (2024). Structured Pruning Graph Neural Networks Models for Inference Acceleration [Master's thesis, Youngstown State University]. OhioLINK Electronic Theses and Dissertations Center. http://rave.ohiolink.edu/etdc/view?acc_num=ysu173375500625166

    APA Style (7th edition)

  • Naghibi Saleh, Gelareh. Structured Pruning Graph Neural Networks Models for Inference Acceleration. 2024. Youngstown State University, Master's thesis. OhioLINK Electronic Theses and Dissertations Center, http://rave.ohiolink.edu/etdc/view?acc_num=ysu173375500625166.

    MLA Style (8th edition)

  • Naghibi Saleh, Gelareh. "Structured Pruning Graph Neural Networks Models for Inference Acceleration." Master's thesis, Youngstown State University, 2024. http://rave.ohiolink.edu/etdc/view?acc_num=ysu173375500625166

    Chicago Manual of Style (17th edition)

ysu173375500625166
68
© 2024, all rights reserved.
This open access ETD is published by Youngstown State University and OhioLINK.