pith. sign in

arxiv: 1909.01315 · v2 · pith:RY4YAJTKnew · submitted 2019-09-03 · 💻 cs.LG · stat.ML

Deep Graph Library: A Graph-Centric, Highly-Performant Package for Graph Neural Networks

Minjie Wang , Da Zheng , Zihao Ye , Quan Gan , Mufei Li , Xiang Song , Jinjing Zhou , Chao Ma
show 7 more authors
Lingfan Yu Yu Gai Tianjun Xiao Tong He George Karypis Jinyang Li Zheng Zhang
This is my paper

Pith reviewed 2026-05-24 00:37 UTC · model grok-4.3

classification 💻 cs.LG stat.ML
keywords graph neural networksdeep graph librarysparse tensor operationsframework neutralperformance evaluationgraph abstractionparallel computation
0
0 comments X

The pith

DGL uses a graph-centric abstraction to optimize sparse tensor operations for GNNs across frameworks.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces Deep Graph Library as a package built around graph structures as the main programming model for graph neural networks. It reduces common GNN patterns to a small set of generalized sparse tensor operations that support heavy parallel execution. The design stays neutral to underlying deep learning frameworks so existing components can be reused without change. If these choices hold, developers could run the same GNN code faster and with lower memory use on multiple backends while keeping overhead low even on modest workloads.

Core claim

DGL distills the computational patterns of GNNs into a few generalized sparse tensor operations suitable for extensive parallelization. By advocating graph as the central programming abstraction, DGL can perform optimizations transparently. By cautiously adopting a framework-neutral design, DGL allows users to easily port and leverage the existing components across multiple deep learning frameworks, delivering significant gains in speed and memory over other GNN-oriented frameworks on a variety of benchmarks with little overhead on small-scale workloads.

What carries the argument

Graph as the central programming abstraction that turns GNN workloads into optimized sparse tensor operations.

If this is right

  • GNN training runs faster and uses less memory on standard benchmarks than with prior frameworks.
  • Code written once can move between deep learning backends without losing the performance gains.
  • Small workloads incur almost no extra cost, allowing the same library to serve both research prototypes and production jobs.
  • Optimizations happen automatically once the graph structure is expressed, reducing the need for manual tuning.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Adoption could shift how researchers express graph algorithms so that hardware and compiler teams target the same sparse primitives.
  • The neutral design might make it easier to combine DGL with new tensor runtimes that appear after the original paper.
  • Future work could test whether the same abstraction yields gains on graph problems outside the GNN setting, such as graph analytics pipelines.

Load-bearing premise

The chosen benchmarks and workloads stand in for the full range of real GNN applications without exposing hidden costs at other scales or usage patterns.

What would settle it

Running the same set of models on a new workload or larger scale where DGL shows no consistent speed or memory advantage over the compared frameworks.

read the original abstract

Advancing research in the emerging field of deep graph learning requires new tools to support tensor computation over graphs. In this paper, we present the design principles and implementation of Deep Graph Library (DGL). DGL distills the computational patterns of GNNs into a few generalized sparse tensor operations suitable for extensive parallelization. By advocating graph as the central programming abstraction, DGL can perform optimizations transparently. By cautiously adopting a framework-neutral design, DGL allows users to easily port and leverage the existing components across multiple deep learning frameworks. Our evaluation shows that DGL significantly outperforms other popular GNN-oriented frameworks in both speed and memory consumption over a variety of benchmarks and has little overhead for small scale workloads.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

0 major / 2 minor

Summary. The paper presents the design principles and implementation of Deep Graph Library (DGL), a package for graph neural networks. It distills GNN computational patterns into generalized sparse tensor operations, advocates graph as the central programming abstraction for transparent optimizations, adopts a framework-neutral design to allow porting across deep learning frameworks, and reports empirical results showing significant outperformance in speed and memory consumption over other GNN frameworks across benchmarks with little overhead for small-scale workloads.

Significance. If the performance results hold under scrutiny of the full evaluation, DGL would offer a useful systems contribution to the GNN community by providing an efficient, graph-centric abstraction with cross-framework compatibility. The framework-neutral design and emphasis on transparent optimizations are notable strengths that could aid adoption.

minor comments (2)
  1. [Abstract] Abstract: the claim of outperformance 'over a variety of benchmarks' would be strengthened by briefly naming the compared frameworks, models, and datasets (or directing to the evaluation section) so readers can immediately gauge scope.
  2. The manuscript would benefit from an explicit discussion of potential hidden costs or usage patterns where the framework-neutral design might incur overhead, to address the representativeness of the reported workloads.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the positive summary, significance assessment, and recommendation of minor revision. We appreciate the recognition of DGL's graph-centric design, transparent optimizations, and cross-framework compatibility as notable strengths.

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper is a systems description of the DGL library design, implementation, and empirical performance evaluation rather than a mathematical derivation. Performance claims rest on benchmark measurements (speed, memory) across workloads, which are external empirical results and not quantities derived from fitted parameters, self-referential equations, or self-citation chains within the paper. No load-bearing steps reduce by construction to inputs; the framework-neutral design and graph-centric abstractions are presented as engineering choices supported by implementation details and measurements. This is the expected finding for a non-derivational systems paper.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

This is a systems and software-engineering paper whose central claim is an empirical performance comparison of a library implementation. No mathematical free parameters, domain axioms, or invented physical entities are required or introduced.

pith-pipeline@v0.9.0 · 5694 in / 1140 out tokens · 17415 ms · 2026-05-24T00:37:02.614378+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

  • Cost.FunctionalEquation washburn_uniqueness_aczel unclear
    ?
    unclear

    Relation between the paper passage and the cited Recognition theorem.

    DGL distills the computational patterns of GNNs into a few generalized sparse tensor operations suitable for extensive parallelization. By advocating graph as the central programming abstraction...

  • Foundation.DimensionForcing alexander_duality_circle_linking unclear
    ?
    unclear

    Relation between the paper passage and the cited Recognition theorem.

    Our evaluation shows that DGL significantly outperforms other popular GNN-oriented frameworks in both speed and memory consumption...

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Forward citations

Cited by 23 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Ocean: Fast Estimation-Based Sparse General Matrix-Matrix Multiplication on GPU

    cs.DC 2026-04 unverdicted novelty 7.0

    Ocean uses HyperLogLog estimators to skip the costly symbolic phase of GPU SpGEMM, pairs it with dynamic workflow choice and a shared-plus-global hash accumulator, and reports 1.4-2.8x speedups over prior GPU implementations.

  2. AsyncSparse: Accelerating Sparse Matrix-Matrix Multiplication on Asynchronous GPU Architectures

    cs.DC 2026-04 unverdicted novelty 7.0

    AsyncSparse presents BCSR and WCSR kernels that use TMA and warp specialization to accelerate SpMM, outperforming prior libraries by 1.47-6.24x on SuiteSparse and achieving 2.66x end-to-end speedup on Qwen2.5-7B at 90...

  3. How Hard Is It for Message-Passing GNNs to Simulate One Weisfeiler-Lehman Color-Refinement Step?

    cs.LG 2024-10 unverdicted novelty 7.0

    Oblivious MPGNNs cannot simulate WL color refinement with shallow depth and small messages without randomness; bounded-error randomness enables logarithmic resources for large color sets, while small color sets force ...

  4. How Attentive are Graph Attention Networks?

    cs.LG 2021-05 conditional novelty 7.0

    GAT uses static attention where neighbor rankings ignore the query node and thus cannot express some graph problems; GATv2 enables dynamic attention and outperforms GAT on 11 OGB and other benchmarks with equal parameters.

  5. Robust Multimodal Recommendation via Graph Retrieval-Enhanced Modality Completion

    cs.IR 2026-05 unverdicted novelty 6.0

    GRE-MC retrieves relevant subgraphs and uses a graph transformer plus sparse codebook to complete missing modalities, outperforming prior methods on recommendation benchmarks.

  6. LogosKG: Hardware-Optimized Scalable and Interpretable Knowledge Graph Retrieval

    cs.CL 2026-04 unverdicted novelty 6.0

    LogosKG delivers a novel hardware-aligned system for efficient multi-hop retrieval on billion-edge knowledge graphs without sacrificing fidelity, demonstrated via biomedical KG-LLM applications.

  7. Scalable and Adaptive Parallel Training of Graph Transformer on Large Graphs

    cs.DC 2026-04 unverdicted novelty 6.0

    A new distributed framework for graph transformer training auto-selects parallel strategies and optimizes sparse operations to deliver up to 6x speedup on 8 GPUs and 78% memory reduction.

  8. Modern Structure-Aware Simplicial Spatiotemporal Neural Network

    cs.LG 2026-04 unverdicted novelty 6.0

    ModernSASST is the first simplicial complex-based spatiotemporal model that combines random walks on high-dimensional complexes with parallelizable temporal convolutional networks for efficient high-order topology capture.

  9. Cluster Attention for Graph Machine Learning

    cs.LG 2026-04 unverdicted novelty 6.0

    Cluster attention uses off-the-shelf community detection to define attention scopes within graph clusters, augmenting MPNNs and Graph Transformers to achieve larger receptive fields with preserved structural inductive...

  10. Communication-free Sampling and 4D Hybrid Parallelism for Scalable Mini-batch GNN Training

    cs.LG 2026-04 unverdicted novelty 6.0

    ScaleGNN uses communication-free sampling and 4D parallelism to scale mini-batch GNN training to 2048 GPUs, achieving 3.5x speedup over prior state-of-the-art on ogbn-products.

  11. FlexMS is a flexible framework for benchmarking deep learning-based mass spectrum prediction tools in metabolomics

    cs.AI 2026-02 unverdicted novelty 6.0

    FlexMS is a new flexible benchmarking framework that lets researchers dynamically combine deep learning architectures and evaluate their mass spectrum prediction performance on public metabolomics datasets using multi...

  12. SHIRO: Near-Optimal Communication Strategies for Distributed Sparse Matrix Multiplication

    cs.DC 2025-12 unverdicted novelty 6.0

    SHIRO achieves geometric mean speedups of 221.5x to 8.8x over four baselines in distributed SpMM on up to 128 GPUs by exploiting sparsity patterns and two-tier network topologies.

  13. Torch Geometric Pool: the PyTorch library for pooling in Graph Neural Networks

    cs.LG 2025-12 accept novelty 6.0

    A new open-source library standardizes 20 hierarchical graph pooling operations under one SRCL interface with uniform outputs and batch handling for PyTorch Geometric.

  14. Detecting LLM-Generated Spam Reviews by Integrating Language Model Embeddings and Graph Neural Network

    cs.CL 2025-10 unverdicted novelty 6.0

    Introduces FraudSquad, a hybrid model using language model embeddings and a gated graph transformer that outperforms baselines on newly created LLM-generated spam review datasets.

  15. Modal Decomposition and Identification for a Population of Structures Using Physics-Informed Graph Neural Networks and Transformers

    cs.CE 2025-05 unverdicted novelty 6.0

    A physics-informed GNN-transformer model performs unsupervised modal decomposition and identification for populations of structures from sparse dynamic measurements.

  16. Pretrained Event Classification Model for High Energy Physics Analysis

    hep-ph 2024-12 unverdicted novelty 6.0

    A GNN pretrained on 120M simulated HEP events generalizes to unseen processes and ATLAS data; fine-tuning boosts accuracy especially with small datasets, with CKA showing preserved encoders but altered intermediate layers.

  17. Learning Spatial-Preserving Hierarchical Representations for Digital Pathology

    cs.CV 2024-06 unverdicted novelty 6.0

    SPAN is a hierarchical attention framework that constructs multi-scale pyramid representations from single-scale patch inputs for WSI classification and segmentation while preserving spatial relationships.

  18. Graph Transductive Sharpening: Leveraging Unlabeled Predictions in Node Classification

    cs.LG 2026-05 unverdicted novelty 5.0

    Transductive Sharpening adds an entropy-minimization term on unlabeled-node predictions to the training objective for graph node classification.

  19. GreenDyGNN: Runtime-Adaptive Energy-Efficient Communication for Distributed GNN Training

    cs.DC 2026-04 unverdicted novelty 5.0

    GreenDyGNN applies Double-DQN to adapt cache management in distributed GNN training, cutting energy by up to 43% under congestion versus static policies.

  20. TabEmb: Joint Semantic-Structure Embedding for Table Annotation

    cs.LG 2026-04 unverdicted novelty 5.0

    TabEmb decouples LLM-based semantic column embeddings from graph-based structural modeling to produce joint representations that improve table annotation tasks.

  21. AutoGraphAD: Unsupervised network anomaly detection using Variational Graph Autoencoders

    cs.CR 2025-11 unverdicted novelty 5.0

    AutoGraphAD applies a heterogeneous variational graph autoencoder with unsupervised and contrastive learning to detect network anomalies on connection-IP graphs without labeled data, achieving comparable performance t...

  22. G-reasoner: Foundation Models for Unified Reasoning over Graph-structured Knowledge

    cs.AI 2025-09 unverdicted novelty 5.0

    G-reasoner uses QuadGraph abstraction and a 34M-parameter graph foundation model integrated with LLMs to enable scalable reasoning over diverse graph-structured knowledge, outperforming baselines on six benchmarks.

  23. Software and computing for Run 3 of the ATLAS experiment at the LHC

    hep-ex 2024-04 unverdicted novelty 2.0

    ATLAS reports on its Run 3 software infrastructure for data management, workflows, databases, validation, and physics analysis tools at the LHC.

Reference graph

Works this paper leans on

17 extracted references · 17 canonical work pages · cited by 23 Pith papers · 10 internal anchors

  1. [1]

    Relational inductive biases, deep learning, and graph networks

    Peter W Battaglia, Jessica B Hamrick, Victor Bapst, Alvaro Sanchez-Gonzalez, Vinicius Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo, Adam Santoro, Ryan Faulkner, et al. Relational inductive biases, deep learning, and graph networks. arXiv preprint arXiv:1806.01261,

    work page internal anchor Pith review Pith/arXiv arXiv

  2. [2]

    Efficient sparse matrix-vector multiplication on cuda

    Nathan Bell and Michael Garland. Efficient sparse matrix-vector multiplication on cuda. Technical report, Nvidia Technical Report NVR-2008-004, Nvidia Corporation,

    work page 2008

  3. [3]

    Graph Convolutional Matrix Completion

    Rianne van den Berg, Thomas N Kipf, and Max Welling. Graph convolutional matrix completion. arXiv preprint arXiv:1706.02263,

    work page internal anchor Pith review Pith/arXiv arXiv

  4. [4]

    Spectral Networks and Locally Connected Networks on Graphs

    Joan Bruna, Wojciech Zaremba, Arthur Szlam, and Yann LeCun. Spectral networks and locally connected networks on graphs. arXiv preprint arXiv:1312.6203,

    work page internal anchor Pith review Pith/arXiv arXiv

  5. [5]

    MXNet: A Flexible and Efficient Machine Learning Library for Heterogeneous Distributed Systems

    Tianqi Chen, Mu Li, Yutian Li, Min Lin, Naiyan Wang, Minjie Wang, Tianjun Xiao, Bing Xu, Chiyuan Zhang, and Zheng Zhang. Mxnet: A flexible and efficient machine learning library for heterogeneous distributed systems. arXiv preprint arXiv:1512.01274,

    work page internal anchor Pith review Pith/arXiv arXiv

  6. [6]

    Matthias Fey and Jan E. Lenssen. Fast graph representation learning with PyTorch Geometric. CoRR, abs/1903.02428,

    work page internal anchor Pith review Pith/arXiv arXiv 1903

  7. [7]

    Embedding logical queries on knowledge graphs

    Will Hamilton, Payal Bajaj, Marinka Zitnik, Dan Jurafsky, and Jure Leskovec. Embedding logical queries on knowledge graphs. In Advances in Neural Information Processing Systems, pp. 2030– 2041,

    work page 2030

  8. [8]

    Open graph benchmark: Datasets for machine learning on graphs

    Weihua Hu, Matthias Fey, Marinka Zitnik, Yuxiao Dong, Hongyu Ren, Bowen Liu, Michele Catasta, and Jure Leskovec. Open graph benchmark: Datasets for machine learning on graphs. arXiv preprint arXiv:2005.00687,

    work page arXiv 2005

  9. [9]

    Mathematical founda- tions of the graphblas

    Jeremy Kepner, Peter Aaltonen, David Bader, Aydin Buluç, Franz Franchetti, John Gilbert, Dylan Hutchison, Manoj Kumar, Andrew Lumsdaine, Henning Meyerhenke, et al. Mathematical founda- tions of the graphblas. In 2016 IEEE High Performance Extreme Computing Conference (HPEC), pp. 1–9. IEEE,

    work page 2016

  10. [10]

    Analysis of the Impact of Negative Sampling on Link Prediction in Knowledge Graphs

    Bhushan Kotnis and Vivi Nastase. Analysis of the impact of negative sampling on link prediction in knowledge graphs. arXiv preprint arXiv:1708.06816,

    work page internal anchor Pith review Pith/arXiv arXiv

  11. [11]

    Neugraph: parallel deep neural network computation on large graphs

    Lingxiao Ma, Zhi Yang, Youshan Miao, Jilong Xue, Ming Wu, Lidong Zhou, and Yafei Dai. Neugraph: parallel deep neural network computation on large graphs. In 2019{USENIX} Annual Technical Conference ({USENIX}{ATC} 19), pp. 443–458,

    work page 2019

  12. [12]

    Modeling Relational Data with Graph Convolutional Networks

    Michael Schlichtkrull, Thomas N Kipf, Peter Bloem, Rianne van den Berg, Ivan Titov, and Max Welling. Modeling relational data with graph convolutional networks. arXiv preprint arXiv:1703.06103,

    work page internal anchor Pith review Pith/arXiv arXiv

  13. [13]

    Simplifying Graph Convolutional Networks

    Felix Wu, Tianyi Zhang, Amauri Holanda de Souza Jr, Christopher Fifty, Tao Yu, and Kilian Q Weinberger. Simplifying graph convolutional networks. arXiv preprint arXiv:1902.07153,

    work page internal anchor Pith review Pith/arXiv arXiv 1902

  14. [14]

    How Powerful are Graph Neural Networks?

    Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. How powerful are graph neural networks? arXiv preprint arXiv:1810.00826,

    work page internal anchor Pith review Pith/arXiv arXiv

  15. [15]

    Fast Sparse Matrix-Vector Multiplication on GPUs: Implications for Graph Mining

    Xintian Yang, Srinivasan Parthasarathy, and Ponnuswamy Sadayappan. Fast sparse matrix-vector multiplication on gpus: implications for graph mining. arXiv preprint arXiv:1103.2405,

    work page internal anchor Pith review Pith/arXiv arXiv

  16. [16]

    Aligraph: a comprehensive graph neural network platform

    Rong Zhu, Kun Zhao, Hongxia Yang, Wei Lin, Chang Zhou, Baole Ai, Yong Li, and Jingren Zhou. Aligraph: a comprehensive graph neural network platform. Proceedings of the VLDB Endowment, 12(12):2094–2105,

    work page 2094

  17. [17]

    fan _avg

    and a nearest neighbor graph generated by (Qi et al., 2017), with varying feature sizes. 16 Preprint 1 from torch import nn 2 3 class SAGEConv(nn.Module): 4 def __init__(self, in _feat, out _feat, 5 feat _drop=0., activation=None): 6 pass 7 def forward(self, tensor): 8 pass (a) PyTorch 1 from tensorflow.keras import layers 2 3 class SAGEConv(layers.Layer)...

    work page 2017