Pith. sign in

REVIEW 2 major objections 2 minor 27 references

GreenGNN: Energy-Aware Windowed Communication Optimization for Distributed GNN Training

T0 review · 2 major / 2 minor · reviewed 2026-06-28 · grok-4.3

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read GreenGNN groups mini-batches into windows to merge remote feature fetches and cut distributed GNN energy use by 27-43%.

desk verdict GreenGNN uses windowed bulk transfers and a simulator to pick window size for lower energy in distributed GNN training, but the simulator's accuracy on real multi-epoch runs is not shown. read the letter →

arxiv 2606.02916 v1 pith:UGCAEWDE submitted 2026-06-01 cs.DC

Arefin Niam , Tevfik Kosar , M. S. Q. Zulkar Nine This is my paper · ORCID
classification cs.DC
keywords distributedGNNtrainingenergyoptimizationwindowedcommunicationbulktransfersfeaturecachingdiscrete-eventsimulationremoteprocedurecalls
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper shows that sampling-based GNN training on clusters triggers many small remote feature requests per epoch, each carrying fixed initiation costs and leaving GPUs stalled at high baseline power. GreenGNN counters this by collecting W consecutive mini-batches into a window, staging the hot features for that window in a local cache, and collapsing the remote requests from each partition owner into a few bulk transfers. Window size W is chosen once before training by replaying a one-epoch trace in a discrete-event simulator that scores each candidate W with a hybrid energy model. When the selected W is used on a real four-node GPU cluster, total system energy falls 27-43% and throughput rises as much as 3.9x while an on-demand path still handles any cache misses.

What carries the argument

Window size W, selected offline by a discrete-event simulator that replays a deterministic one-epoch access trace under a hybrid energy model; it trades communication amortization against hot-set staleness.

What would settle it

Run the same workloads on the four-node cluster with the simulator-chosen W and with several neighboring values of W, then measure whether real energy and throughput fall outside the reported 27-43% and 3.9x ranges or whether model accuracy degrades from staleness.

Watch

Extended reading notes

Core claim

GreenGNN reduces total system energy by 27--43% relative to baseline while improving end-to-end throughput by up to 3.9x. GPU energy drops by 36--71%, driven by fewer RPC initiations and lower GPU stall time. The system achieves these gains by grouping training into windows of W consecutive mini-batches, staging each window's hot features in a local cache, and merging remote requests from each partition owner into a small number of bulk transfers, with W selected offline by a discrete-event simulator that replays a deterministic one-epoch access trace under a hybrid energy model.

Load-bearing premise

The simulator that replays a fixed one-epoch trace will pick a window size W whose energy savings appear in live runs without unacceptable staleness or extra cache misses.

Editorial extensions

If this is right

  • Fewer RPC initiations directly reduce fixed protocol and initiation energy per remote feature.
  • Lower GPU stall time reduces the fraction of time the GPUs draw baseline power without useful work.
  • Bulk transfers amortize per-request costs across many features while the on-demand path still serves misses.
  • The offline simulator choice of W remains valid across the tested datasets and batch sizes.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same windowing pattern could be applied to other sampling-based distributed workloads whose remote accesses exhibit short-term temporal locality.
  • Periodic re-simulation on updated access traces could let W adapt during a long multi-epoch run.
  • Lower per-epoch energy could allow clusters to train larger graphs under fixed power or cooling budgets.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 2 minor

Summary. GreenGNN is a distributed GNN training system that groups mini-batches into windows of size W to exploit short-lived temporal locality in neighbor sampling, stages hot features in a local cache, and merges remote feature requests into bulk transfers to amortize RPC overhead. Window size W is chosen offline by a discrete-event simulator that replays a deterministic one-epoch access trace under a hybrid energy model. The system is implemented on DGL and evaluated on a 4-node GPU cluster, where it is claimed to reduce total system energy by 27--43% relative to baseline while improving end-to-end throughput by up to 3.9x, with GPU energy reductions of 36--71% driven by fewer RPC initiations and lower stall time.

Significance. If the simulator's hybrid energy model and single-epoch trace reliably predict hardware behavior across epochs without excess staleness or cache-miss penalties, the windowed bulk-communication approach could offer a practical lever for lowering the energy cost of communication-bound distributed GNN training. The engineering idea of trading controlled staleness for amortized RPC cost is straightforward and potentially impactful given the scale of modern GNN workloads, though the magnitude of the reported gains would need to be shown to generalize beyond the specific 4-node setup.

major comments (2)
  1. [Evaluation] Evaluation section: the headline energy (27--43%) and throughput (up to 3.9x) numbers are produced by first running the discrete-event simulator to select W and then measuring the real DGL implementation with that W; however, the manuscript supplies no direct validation that the simulator's hybrid energy model predictions match measured GPU/RPC power draw on the target hardware for the chosen W values, nor any multi-epoch trace comparison to confirm that single-epoch deterministic replay captures cross-epoch access variation.
  2. [Evaluation] Evaluation section: the reported measurements lack error bars, statistics over multiple runs, or explicit baseline configuration details (exact DGL version, partition sizes, dataset cardinalities, and RPC implementation), making it impossible to assess whether the claimed gains are robust or sensitive to post-hoc choices in the simulator or experimental setup.
minor comments (2)
  1. [Abstract] The abstract states results 'across datasets and batch sizes' but the provided text does not enumerate the concrete datasets, batch sizes, or graph sizes used; this should be stated explicitly in the evaluation section with a table.
  2. Notation for the hybrid energy model components (e.g., fixed RPC cost, baseline GPU power, stall power) is introduced without an equation or parameter table; adding a compact model definition would improve reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the insightful comments on the evaluation methodology. We provide point-by-point responses below and commit to revisions that enhance the clarity and robustness of the experimental results.

read point-by-point responses

  1. Referee: [Evaluation] Evaluation section: the headline energy (27--43%) and throughput (up to 3.9x) numbers are produced by first running the discrete-event simulator to select W and then measuring the real DGL implementation with that W; however, the manuscript supplies no direct validation that the simulator's hybrid energy model predictions match measured GPU/RPC power draw on the target hardware for the chosen W values, nor any multi-epoch trace comparison to confirm that single-epoch deterministic replay captures cross-epoch access variation.

    Authors: The simulator serves as an offline tool to select an appropriate window size W by estimating the energy trade-offs using the hybrid model and a one-epoch trace. The reported performance gains, including energy reductions and throughput improvements, are derived from direct measurements on the 4-node GPU cluster running the actual GreenGNN implementation. We recognize that a direct validation of the simulator's predictions against hardware measurements for the specific W values is not presented in the manuscript. In the revised version, we will include such a validation by comparing the simulator's energy estimates for the selected W with the measured energy consumption from the real experiments. For the multi-epoch aspect, the deterministic replay is based on the fact that the sampling process depends primarily on the fixed graph structure and the sequential mini-batch ordering, leading to highly similar access patterns across epochs. Nevertheless, to address this, we will add a comparison of access traces from two epochs to demonstrate the limited variation. revision: yes

  2. Referee: [Evaluation] Evaluation section: the reported measurements lack error bars, statistics over multiple runs, or explicit baseline configuration details (exact DGL version, partition sizes, dataset cardinalities, and RPC implementation), making it impossible to assess whether the claimed gains are robust or sensitive to post-hoc choices in the simulator or experimental setup.

    Authors: We agree that providing error bars and multiple-run statistics, along with more detailed baseline configurations, would improve the reproducibility and credibility of the results. In the revised manuscript, we will report all key metrics as averages over at least five independent runs, including standard deviations as error bars. Additionally, we will expand the 'Experimental Setup' section to specify the exact DGL version used, the partition sizes for each dataset, the cardinalities of the datasets, and the precise RPC implementation details, including any custom configurations for bulk transfers. revision: yes

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; claims rest on hardware measurements, not self-referential derivations

full rationale

The paper presents GreenGNN as an empirical system: it selects window size W offline via a discrete-event simulator replaying a one-epoch trace under a hybrid energy model, then measures actual energy (27-43% reduction) and throughput (up to 3.9x) on a 4-node DGL cluster. No equations or derivations are shown that reduce a claimed prediction back to the simulator inputs by construction; the reported savings are post-implementation hardware outcomes rather than fitted parameters renamed as predictions. No self-citations, uniqueness theorems, or ansatzes are invoked as load-bearing steps in the provided text. The derivation chain is self-contained against external benchmarks (real runs), warranting a score of 0.

Assumptions & free parameters 1 free parameters · 2 assumptions · 0 invented entities

The central claim rests on the existence of exploitable temporal locality in neighbor sampling and on the accuracy of an offline hybrid energy model simulator; window size W is a tunable parameter whose value is chosen by that simulator.

free parameters (1)
  • W
    Window size that trades off communication amortization against hot-set staleness; selected offline by the simulator.
assumptions (2)
  • domain assumption Neighbor sampling produces bursty, short-lived temporal locality that can be captured by a deterministic one-epoch access trace.
    Invoked to justify staging hot features in local cache and merging requests into bulk transfers.
  • domain assumption The hybrid energy model inside the discrete-event simulator correctly predicts real-system energy and stall time.
    Used to choose W and to claim the reported energy reductions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GreenGNN: Energy-Aware Windowed Communication Optimization for Distributed GNN Training." pith.science (2026). https://pith.science/paper/UGCAEWDE

@misc{pith2026260602916,
  author       = {Pith},
  title        = {Pith review of: GreenGNN: Energy-Aware Windowed Communication Optimization for Distributed GNN Training},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UGCAEWDE}},
  note         = {Machine review of arXiv:2606.02916}
}
read the original abstract

Large-scale graph neural network (GNN) training often requires distributed clusters because graph structure and feature tensors no longer fit in a single node's memory. In sampling-based training, each mini-batch expands into a receptive field that spans partitions and triggers thousands of remote feature fetches per epoch. This wastes energy for two main reasons: each small RPC pays a fixed initiation and protocol cost, and GPUs continue drawing substantial baseline power while waiting for remote features. We present GreenGNN, an energy-aware distributed GNN training system that reduces communication energy by exploiting the bursty, short-lived temporal locality of neighbor sampling. GreenGNN groups training into windows of W consecutive mini-batches, stages each window's hot features in a local cache, and merges remote requests from each partition owner into a small number of bulk transfers. This amortizes RPC overhead across many features while preserving an on-demand path for cache misses. Because window size controls the trade-off between communication amortization and hot-set staleness, GreenGNN selects W offline using a discrete-event simulator that replays a deterministic one-epoch access trace with a hybrid energy model. We implement GreenGNN on DGL and evaluate it on a 4-node GPU cluster with benchmark datasets. Across datasets and batch sizes, GreenGNN reduces total system energy by 27--43% relative to baseline while improving end-to-end throughput by up to 3.9x. GPU energy drops by 36--71%, driven by fewer RPC initiations and lower GPU stall time.

Figures

Figures reproduced from arXiv: 2606.02916 by the authors.

Figure 1
Figure 1. Energy breakdown of distributed GraphSAGE training [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Initiation vs. payload energy as a function of transfer [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. GreenGNN architecture overview. Circled numbers [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: GreenGNN simulator workflow. For each candidate [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: GreenGNN energy profiling pipeline. The Profiler [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 7
Figure 7. Figure 7: Total energy consumption (GPU + CPU, all nodes) [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: GPU energy at B=2000. GreenGNN achieves 37– 71% GPU energy reduction over Default DGL. GraphStorm on Papers100M is truncated (actual: 340 kJ). the energy of GreenGNN. This increase is driven by post￾training overhead (model serialization, distributed inference setup) w…
Figure 11
Figure 11. Figure 11: GreenGNN’s reduction over Default DGL (%) across [PITH_FULL_IMAGE:figures/full_fig_p009_11.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

27 extracted references · 2 canonical work pages

  1. [1]

    Graph convolutional neural networks for web-scale rec- ommender systems,

    R. Ying, R. He, K. Chen, P. Eksombatchai, W. L. Hamilton, and J. Leskovec, “Graph convolutional neural networks for web-scale rec- ommender systems,” inProceedings of the 24th ACM SIGKDD inter- national conference on knowledge discovery & data mining, 2018, pp. 974–983

    2018

  2. [2]

    Graph neural networks for social recommendation,

    W. Fan, Y . Ma, Q. Li, Y . He, E. Zhao, J. Tang, and D. Yin, “Graph neural networks for social recommendation,” inThe world wide web conference, 2019, pp. 417–426

    2019

  3. [3]

    Enhanc- ing graph neural network-based fraud detectors against camouflaged fraudsters,

    Y . Dou, Z. Liu, L. Sun, Y . Deng, H. Peng, and P. S. Yu, “Enhanc- ing graph neural network-based fraud detectors against camouflaged fraudsters,” inProceedings of the 29th ACM international conference on information & knowledge management, 2020, pp. 315–324

    2020

  4. [4]

    Graph neural networks and their current applications in bioinformatics,

    X.-M. Zhang, L. Liang, L. Liu, and M.-J. Tang, “Graph neural networks and their current applications in bioinformatics,”Frontiers in genetics, vol. 12, p. 690049, 2021

    2021

  5. [5]

    A comprehensive survey on graph neural networks,

    Z. Wu, S. Pan, F. Chen, G. Long, C. Zhang, and P. S. Yu, “A comprehensive survey on graph neural networks,”IEEE transactions on neural networks and learning systems, vol. 32, no. 1, pp. 4–24, 2020

    2020

  6. [6]

    Inductive representation learning on large graphs,

    W. Hamilton, Z. Ying, and J. Leskovec, “Inductive representation learning on large graphs,”Advances in neural information processing systems, vol. 30, 2017

    2017

  7. [7]

    Snap: A general-purpose network analysis and graph-mining library,

    J. Leskovec and R. Sosi ˇc, “Snap: A general-purpose network analysis and graph-mining library,”ACM Transactions on Intelligent Systems and Technology (TIST), vol. 8, no. 1, pp. 1–20, 2016

    2016

  8. [8]

    One trillion edges: Graph processing at facebook-scale,

    A. Ching, S. Edunov, M. Kabiljo, D. Logothetis, and S. Muthukrishnan, “One trillion edges: Graph processing at facebook-scale,”Proceedings of the VLDB Endowment, vol. 8, no. 12, pp. 1804–1815, 2015

    2015

Show all 27 references
  1. [9]

    Parallel and distributed graph neural networks: An in-depth concurrency analysis,

    M. Besta and T. Hoefler, “Parallel and distributed graph neural networks: An in-depth concurrency analysis,”IEEE Transactions on Pattern Anal- ysis and Machine Intelligence, vol. 46, no. 5, pp. 2584–2606, 2024

    2024

  2. [10]

    Distdgl: Distributed graph neural network training for billion-scale graphs,

    D. Zheng, C. Ma, M. Wang, J. Zhou, Q. Su, X. Song, Q. Gan, Z. Zhang, and G. Karypis, “Distdgl: Distributed graph neural network training for billion-scale graphs,” in2020 IEEE/ACM 10th Workshop on Irregular Applications: Architectures and Algorithms (IA3). IEEE, 2020, pp. 36–44

    2020

  3. [11]

    P3: Distributed deep graph learning at scale,

    S. Gandhi and A. P. Iyer, “P3: Distributed deep graph learning at scale,” in15th{USENIX}Symposium on Operating Systems Design and Implementation ({OSDI}21), 2021, pp. 551–568

    2021

  4. [12]

    Dgs: Communication-efficient graph sampling for distributed gnn training,

    X. Wan, K. Chen, and Y . Zhang, “Dgs: Communication-efficient graph sampling for distributed gnn training,” in2022 IEEE 30th International Conference on Network Protocols (ICNP). IEEE, 2022, pp. 1–11

    2022

  5. [13]

    Distributed graph neural network training: A survey,

    Y . Shao, H. Li, X. Gu, H. Yin, Y . Li, X. Miao, W. Zhang, B. Cui, and L. Chen, “Distributed graph neural network training: A survey,”ACM Computing Surveys, vol. 56, no. 8, pp. 1–39, 2024

    2024

  6. [14]

    {BGL}:{GPU-Efficient}{GNN}training by optimizing graph data{I/O}and preprocessing,

    T. Liu, Y . Chen, D. Li, C. Wu, Y . Zhu, J. He, Y . Peng, H. Chen, H. Chen, and C. Guo, “{BGL}:{GPU-Efficient}{GNN}training by optimizing graph data{I/O}and preprocessing,” in20th USENIX Symposium on Networked Systems Design and Implementation (NSDI 23), 2023, pp. 103–118

    2023

  7. [15]

    Legion: Automatically pushing the envelope of{Multi-GPU}system for{Billion-Scale}{GNN}training,

    J. Sun, L. Su, Z. Shi, W. Shen, Z. Wang, L. Wang, J. Zhang, Y . Li, W. Yu, J. Zhouet al., “Legion: Automatically pushing the envelope of{Multi-GPU}system for{Billion-Scale}{GNN}training,” in2023 USENIX Annual Technical Conference (USENIX ATC 23), 2023, pp. 165–179

    2023

  8. [16]

    A fast and high quality multilevel scheme for partitioning irregular graphs,

    G. Karypis and V . Kumar, “A fast and high quality multilevel scheme for partitioning irregular graphs,”SIAM Journal on scientific Computing, vol. 20, no. 1, pp. 359–392, 1998

    1998

  9. [17]

    Pagraph: Scaling gnn training on large graphs via computation-aware caching,

    Z. Lin, C. Li, Y . Miao, Y . Liu, and Y . Xu, “Pagraph: Scaling gnn training on large graphs via computation-aware caching,” inProceedings of the 11th ACM Symposium on Cloud Computing, 2020, pp. 401–415

    2020

  10. [18]

    Gnnlab: a factored system for sample-based gnn training over gpus,

    J. Yang, D. Tang, X. Song, L. Wang, Q. Yin, R. Chen, W. Yu, and J. Zhou, “Gnnlab: a factored system for sample-based gnn training over gpus,” inProceedings of the Seventeenth European Conference on Computer Systems, 2022, pp. 417–434

    2022

  11. [19]

    Four degrees of separation,

    L. Backstrom, P. Boldi, M. Rosa, J. Ugander, and S. Vigna, “Four degrees of separation,” inProceedings of the 4th annual ACM Web science conference, 2012, pp. 33–42

    2012

  12. [20]

    Two-level graph caching for expediting distributed gnn training,

    Z. Zhang, Z. Luo, and C. Wu, “Two-level graph caching for expediting distributed gnn training,” inIEEE INFOCOM 2023-IEEE Conference on Computer Communications. IEEE, 2023, pp. 1–10

    2023

  13. [21]

    H100 tensor core gpu architecture overview,

    N. NVIDIA, “H100 tensor core gpu architecture overview,” 2022

    2022

  14. [22]

    Dgcl: An efficient communication library for distributed gnn training,

    Z. Cai, X. Yan, Y . Wu, K. Ma, J. Cheng, and F. Yu, “Dgcl: An efficient communication library for distributed gnn training,” inProceedings of the Sixteenth European Conference on Computer Systems, 2021, pp. 130–144

    2021

  15. [23]

    Granndis: Fast distributed graph neural network training framework for multi- server clusters,

    J. Song, H. Jang, H. Lim, J. Jung, Y . Kim, and J. Lee, “Granndis: Fast distributed graph neural network training framework for multi- server clusters,” inProceedings of the 2024 International Conference on Parallel Architectures and Compilation Techniques, 2024, pp. 91– 107

    2024

  16. [24]

    Pipegcn: Efficient full-graph training of graph convolu- tional networks with pipelined feature communication,

    C. Wan, Y . Li, C. R. Wolfe, A. Kyrillidis, N. S. Kim, and Y . Lin, “Pipegcn: Efficient full-graph training of graph convolu- tional networks with pipelined feature communication,”arXiv preprint arXiv:2203.10428, 2022

    2022

  17. [25]

    Xgnn: Boosting multi-gpu gnn training via global gnn memory store,

    D. Tang, J. Wang, R. Chen, L. Wang, W. Yu, J. Zhou, and K. Li, “Xgnn: Boosting multi-gpu gnn training via global gnn memory store,” Proceedings of the VLDB Endowment, vol. 17, no. 5, pp. 1105–1118, 2024

    2024

  18. [26]

    Rapidgnn: Communication efficient large- scale distributed training of graph neural networks,

    A. Niam and M. Nine, “Rapidgnn: Communication efficient large- scale distributed training of graph neural networks,”arXiv preprint arXiv:2505.10806, 2025

    2025

  19. [27]

    Graphstorm: all-in-one graph machine learning framework for industry applications,

    D. Zheng, X. Song, Q. Zhu, J. Zhang, T. Vasiloudis, R. Ma, H. Zhang, Z. Wang, S. Adeshina, I. Nisaet al., “Graphstorm: all-in-one graph machine learning framework for industry applications,” inProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining...

    2024

Pith tools

Reviewed June 28, 2026 · model on record in the stance chip above.