Pith. sign in

REVIEW 3 major objections 6 minor 39 references

Round-robin MoE dispatch causes exponential incast; normalizing the traffic matrix prevents it by construction.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 00:07 UTC pith:T7TYALTO

load-bearing objection The exponential incast model is the real contribution; the rate-normalization fix is sound but simple, and the paper overstates its incast-free guarantee outside the exact fluid model. the 3 major comments →

arxiv 2607.26340 v1 pith:T7TYALTO submitted 2026-07-28 cs.NI cs.DCcs.LG

Incast-Free MoE Rate-Based Scheduling

Evyatar Cohen , Jose Yallouz , Alexander Shpiner , Mark Silberstein , Sylvia Ratnasamy , Isaac Keslassy This is my paper
classification cs.NI cs.DCcs.LG
keywords Mixture of Expertsround-robin schedulingincastZipfian trafficrate normalizationtraffic matrixNIC pacingcollective communication time
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper claims that round-robin scheduling of Mixture-of-Experts token dispatch creates a previously unidentified exponential incast: as light flows to unpopular experts drain, all senders converge on the hottest receiver, driving its arrival rate from one link's worth to N times line rate, and leaving nearly the entire workload stuck in switch buffers. It derives a fluid model showing the number of active destinations decays exponentially and the total network backlog approaches the full workload volume at large scale. The proposed fix is a static rate allocation: normalize the traffic demand matrix by the maximum row or column sum, then have each NIC pace each flow at its normalized rate, which by construction keeps every port below line rate. The paper argues that this proactive pacing eliminates incast and lets reactive congestion control be simplified, and supports it with simulations of real and synthetic MoE workloads showing lower collective completion times than round-robin with congestion control.

Core claim

Under round-robin scheduling with Zipfian MoE traffic, the number of destinations a sender still services decays as k(t)=N^{1-t/T}, so the aggregate arrival rate at the most popular receiver grows as λ1(t)=N^{t/T} and peaks at N at the end of the epoch. This creates a lower bound on total network backlog of roughly NT(1-(ln ln N+1)/ln N), which tends to NT at large N—meaning nearly the entire workload volume becomes clogged inside the fabric despite sufficient bisection bandwidth. The paper then shows that dividing the demand matrix by M, the maximum of the worst-case row sum and worst-case column sum, yields rates R_{ij}=D_{ij}/M such that every sender's egress and every receiver's ingress

What carries the argument

The central mechanism is the round-robin queue-depletion boundary: under RR each flow drains at rate 1/k(τ), so the boundary of remaining active flows satisfies an integral equation that yields exponential decay k(t)=N e^{-α t/T}. This feeds the hot-receiver arrival rate λ1(t)=N^{1-?} growth. The proposed fix is the normalization R=D/M with M=max(max_i Σ_j D_{ij}, max_j Σ_i D_{ij}), a simple row/column bound that guarantees every normalized row and column sum≤1. The work also relies on a programmable-NIC rate-limiting interface to enforce these rates with sub-10μs updates.

Load-bearing premise

The incast-free guarantee depends on knowing the complete global traffic matrix before each dispatch epoch and on each NIC being able to reprogram its per-flow rates precisely within tens of microseconds; the paper does not measure the latency of the global gather or the accuracy of rate enforcement.

What would settle it

Instrument an MoE dispatch under round-robin and measure the hot receiver's arrival rate over time: if it does not follow the N^{t/T}-like rise (or if the growth is merely linear), the exponential-incast claim fails. Alternatively, run the proposed rate-based scheduler with a stale or partially known traffic matrix and check whether incast reappears when the matrix drifts during the epoch.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If per-flow rates are enforced as specified, the fabric is provably collision-free during the dispatch epoch, so reactive congestion control no longer needs to throttle and link utilization stays near 100%.
  • MoE collective completion time is reduced relative to round-robin with reactive congestion control, with the gap growing under stronger skewness.
  • Switch buffers can be provisioned for normalized demand rather than worst-case incast, potentially lowering silicon cost and power in future high-speed fabrics.
  • The normalization automatically adapts to asymmetric clusters, since M is set by the slowest bottleneck port, so mixed-speed NICs can coexist without incast.

Where Pith is reading between the lines

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

  • The same normalization applies to any synchronized all-to-all dispatch with a known demand matrix (e.g., gradient all-reduce or TopK token routing), not just MoE, so the result may generalize to a broader class of collective communication workloads.
  • A testable extension is to perturb the demand matrix mid-epoch: the paper's resilience to initial jitter may not cover matrix changes during the dispatch, so measuring how often rate tables must be refreshed under dynamic routing would clarify real-world applicability.
  • If a NIC replaced strict round-robin with weighted fair queueing that accounts for flow sizes, the synchronization might break and the exponential incast could vanish; comparing that scheduler against the proposed rate-based one would reveal whether the benefit comes from pacing or from weighing.

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

3 major / 6 minor

Summary. The paper studies all-to-all MoE token dispatch in datacenter networks. It claims that conventional round-robin (RR) scheduling with Zipfian-skewed MoE traffic produces a previously undiscovered 'exponential incast': as light flows finish, senders synchronously converge on hot receivers, making the hot receiver's aggregate arrival rate grow exponentially. It derives a fluid ODE for the number of active destinations k(t), the resulting arrival rate, and a lower bound on total backlog. It then proposes a proactive rate-allocation framework: normalize the demand matrix D by M, the maximum row or column sum, so every sender/receiver port has ingress rate at most 1. The framework is evaluated in htsim simulations against a round-robin plus UEC NSCC baseline on real Qwen traces and synthetic Zipfian workloads, reporting lower CCT and higher link utilization. An implementation sketch based on NVIDIA DOCA congestion control is outlined as future work.

Significance. If the results hold, the paper makes a useful conceptual contribution: it gives a simple, parameter-free fluid explanation of a practically important MoE incast failure mode, and it proposes an elegant sufficient condition for port-level oversubscription avoidance. Strengths of the manuscript include an explicit ODE derivation with no fitted constants, a clean normalization construction, and evaluation with both real and synthetic workloads against a credible baseline. The main weaknesses are that the exponential incast derivation is specific to Zipfian exponent s=1 while the synthetic evaluation uses s=0.25 and s=0.5, and that the 'incast-free by construction' guarantee relies on exact global knowledge of D and exact hardware rate enforcement, neither of which is demonstrated. These issues are substantial but addressable; I do not see a fatal internal inconsistency in the mathematical core.

major comments (3)
  1. [§3 vs. §6] The derivation of k(t)=N exp(-alpha t/T), and hence the exponential incast law lambda_1(t)=N^{t/T}, is explicitly for the Zipfian exponent s=1 in Eq. (1). For s≠1, the same fluid argument gives dk/dt = -(H_N^{(s)}/(sT)) k^s, which yields polynomial, not exponential, decay. Yet §6 evaluates synthetic workloads with s=0.25 and s=0.5, i.e., outside the regime analyzed in §3. Figure 1 uses one real 128×128 trace, but its effective skewness exponent is not reported. This means the paper's central qualitative claim, 'exponential incast', is not validated for the workloads used in the main evaluation. Please either extend the fluid analysis to general s, or restrict the exponential claim to s=1 and evaluate that case, or report the measured skewness of the real trace.
  2. [§4 and §5] The guarantee that R=D/M eliminates incast depends on two unstated conditions: (i) D is the exact, complete traffic matrix for the epoch and is available to every sender before dispatch; (ii) every NIC enforces each rate R_ij exactly. Section 5 concedes that implementation is future work, and §2.3's assertion that the full matrix can be propagated 'with minimal additional control-plane overhead' is not supported by measurement. Moreover, finite-precision rate limiting can break the column-sum bound: if each R_ij is quantized to the nearest 1/Q, a column sum can exceed 1 by up to N/(2Q); for N=512, Q=64 this is an excess of 4, i.e., a 5x oversubscribed column. The paper provides no quantization analysis or safety margin. Please analyze these implementation errors, propose a corrective normalization, or explicitly state that the guarantee is for an idealized fluid model only.
  3. [§6] The evaluation supports the claims about CCT and link utilization, but it never directly reports the central metric implied by the title: incast. There are no per-scenario maximum queue-occupancy or packet-drop counts for the proposed scheduler versus the RR-NSCC baseline. Since the paper's headline claim is that the framework 'eliminates incast', please add a direct incast metric (e.g., maximum queue depth at receiver ports, number of dropped packets) across the evaluated matrices. This would also clarify whether the CCT gains come from incast avoidance or from other properties of rate pacing.
minor comments (6)
  1. [Abstract] Typo: 'previous-undiscovered' should be 'previously undiscovered'.
  2. [§5] Acronym inconsistency: the section title and body mention DOCA Congestion Control (DCC), while the introduction and abstract refer to DOCA Programmable Congestion Control (PCC). Use one name consistently.
  3. [§3] The sentence 'Under RR, nearly 100% of the entire workload volume becomes completely clogged' is an asymptotic statement. For the evaluated scales the formula gives values well below 100% (e.g., about 55% for N=512). Please qualify the statement as an asymptotic limit rather than a finite-N prediction.
  4. [Fig. 1] The axes are not labeled with units. Please state whether queue size is in packets or bytes, and clarify the time axis (simulated microseconds vs. normalized epoch time).
  5. [§6] The synthetic workloads are described as 'parameterized directly from the DeepSeek-V3 architecture', but the relationship between DeepSeek-V3's top-1/2/4/8 routing and the Zipfian s=0.25/0.5 parameterization is not explained. Please clarify how the Zipfian model maps to the cited architecture.
  6. [§2.3] The claim that the full global traffic matrix can be propagated with 'minimal additional control-plane overhead' is a key practical assumption, but no reference or measurement is provided. Please support this with data or soften the claim.

Circularity Check

0 steps flagged

No significant circularity: the RR incast derivation is a parameter-free fluid model; the scheduler guarantee is explicitly by construction and the only self-citation is non-load-bearing.

full rationale

The central theoretical claim (§3) is a self-contained fluid derivation. Starting from the stated Zipfian demand D_ij = T/(j α) with α≈ln N and the round-robin service-rate assumption that each active flow receives 1/k(t), the paper solves dk/dt = -αk/T to obtain k(t)=N^{1-t/T} and λ1(t)=N^{t/T}; the backlog bound Q_total(T)≈NT(1-(ln ln N+1)/ln N) is a direct integral of these rates. No fitted parameters, no use of the proposed scheduler R, and no author-self citation appear in this chain, so the exponential-incast result is not circular. The proposed scheduler (§4) defines M = max(row sums, column sums) and sets R=D/M; it then states that row and column sums of R are ≤1 'by construction' and that enforcing R_ij prevents port oversubscription. This is an explicit construction guarantee, not a hidden fitted prediction: the paper does not claim to empirically discover the no-incast property, it defines the normalization to achieve it. The subsequent simulations are therefore a consistency check of the implementation rather than an independent test of a predicted phenomenon, which is at most a mild evaluation limitation and not a circular derivation. The only author-overlapping citation is [15] (Gerstein, Silberstein, Keslassy) in §6, used to credit an existing htsim implementation of transport and load-balancing protocols; it is not load-bearing for the paper's claims. No uniqueness theorem, ansatz, or renamed known result is imported from the authors' prior work. Concerns about exact global matrix availability and NIC rate-enforcement precision are implementation assumptions (correctness risks), not circularity.

Axiom & Free-Parameter Ledger

1 free parameters · 5 axioms · 0 invented entities

The paper introduces no new physical entities. Its dependence is on modeling assumptions (identical rows, fair round-robin, global matrix knowledge, sub-10 μs rate programming) and on one hand-picked skewness parameter that drives the exponential claim. The rate-normalization scheduler itself is a design rule, not an entity.

free parameters (1)
  • Zipfian skewness index s = s=1 (theory); s=0.25 and 0.5 (simulations)
    The exponential-decay derivation is solved only for s=1; the simulations use s=0.25 and s=0.5, which would give algebraic rather than exponential decay under the same fluid equations. The choice of s is hand-picked, and the theory-to-experiment link depends on this parameter.
axioms (5)
  • domain assumption All N senders have the identical Zipfian demand vector and are synchronized at epoch start.
    The fluid derivation uses a single boundary k(t) for all senders, which presupposes row-identical demand and synchronous starts. This is stated in §3 but not flagged as a simplification for the real traces.
  • domain assumption Round-robin scheduling gives each active flow exactly 1/k of the sender's link rate.
    The entire differential-equation model rests on this idealization of NIC queue scheduling. Real NICs may have quantization, jitter, or host-side interference.
  • ad hoc to paper The full global traffic matrix D is available to all senders (or a control point) before each MoE epoch.
    The proposed scheduler requires D to compute M and the rates. The paper asserts in §2.3 that this is available with minimal overhead, but it is not measured and is central to the scheme's feasibility.
  • ad hoc to paper Per-flow rates can be enforced and updated in hardware in under 10 μs.
    The 10 μs claim in §5 is based on citations and 'expected' behavior, not on a measured pipeline; the NIC implementation is explicitly future work.
  • domain assumption The network is a non-blocking high-radix fat tree where edge ports are the only bottlenecks.
    The evaluation and the incast analysis assume that core bandwidth is sufficient and that receiver downlink capacity is the limiting resource, stated in §6 topology.

pith-pipeline@v1.3.0-alltime-deepseek · 9741 in / 22841 out tokens · 222782 ms · 2026-08-01T00:07:18.986926+00:00 · methodology

0 comments
read the original abstract

Mixture of Experts (MoE) architectures have become key to large language models; however, their typical round-robin (RR) scheduling introduces significant bottlenecks. In this paper, we demonstrate that RR causes a previously-undiscovered exponential incast phenomenon with MoE traffic. We propose an alternative proactive fair scheduling framework tailored for MoE workloads, which effectively prevents fabric oversubscription. We also outline how it can be implemented in NICs. Finally, through extensive simulations with real and synthetic workloads, we demonstrate that this framework consistently eliminates incast, maintains a near-100% link utilization, and reduces Collective Completion Time (CCT).

Figures

Figures reproduced from arXiv: 2607.26340 by Alexander Shpiner, Evyatar Cohen, Isaac Keslassy, Jose Yallouz, Mark Silberstein, Sylvia Ratnasamy.

Figure 1
Figure 1. Figure 1: Simulation results showing the queue size at [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of the CCT inflation relative to a [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: CCT inflation compared to an ideal lower bound (constant 100% link utilization) under heavily skewed [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The queue-utilization trade-off: reactive [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

39 extracted references · 4 linked inside Pith

  1. [1]

    [n. d.]. Megatron-DeepSpeed. https://github.com/microsoft/Megatron- DeepSpeed

  2. [2]

    https://github.com/smcclure20/csg-htsim

    2025.htsim fork. https://github.com/smcclure20/csg-htsim

    2025

  3. [3]

    NVIDIA Collective Communications Library (NCCL)

    2026. NVIDIA Collective Communications Library (NCCL). https: //developer.nvidia.com/nccl

    2026

  4. [4]

    ROCm Communication Collectives Library (RCCL)

    2026. ROCm Communication Collectives Library (RCCL). https: //github.com/ROCm/rccl

    2026

  5. [5]

    Ultra Ethernet Consortium

    2026. Ultra Ethernet Consortium. https://ultraethernet.org/

    2026

  6. [6]

    Hamdy Abdelkhalik, Yehia Arafa, Nandakishore Santhi, and Abdel- Hameed Badawy. 2022. Demystifying the Nvidia Ampere Architecture through Microbenchmarking and Instruction-level Analysis. InarXiv preprint arXiv:2208.11174

    Pith/arXiv arXiv 2022

  7. [7]

    Abts and J

    D. Abts and J. Kim. 2022.High Performance Datacenter Networks. Springer International Publishing

    2022

  8. [8]

    Garrett Birkhoff. 1946. Three observations on linear algebra.Uni- versidad Nacional de Tucumán. Revista. Serie A: Matemáticas y Física Teórica5 (1946), 147–151

    1946

  9. [9]

    Bonato, A

    T. Bonato, A. Kabbani, A. Ghalayini, M. Papamichael, M. Dohadwala, L. Gianinazzi, M. Khalilov, E. Achermann, D. D. Sensi, and T. Hoefler. 2025. REPS: Recycled entropy packet spraying for adaptive load balancing and failure mitigation. arXiv:2407.21625 [cs.NI] https://arxiv.org/abs/2407 .21625

    arXiv 2025

  10. [10]

    ByteDance / Volcengine. 2026. veScale: Volcengine e-commerce Scale LLM Training Framework. https://github.com/volcengine/veScale. Accessed: July 2026

    2026

  11. [11]

    Ultra Ethernet Consortium. 2026. Ultra Ethernet Specification v1.0.2. https://ultraethernet.org/wp-content/uploads/sites/20/2026/01/UE- Specification-1.0.2-1.pdf

    2026

  12. [12]

    DeepSeek-AI. 2026. DeepEP: High-Performance MoE Communication Library. https://github.com/deepseek-ai/DeepEP

    2026

  13. [13]

    Dixit, P

    A. Dixit, P. Prakash, Y. C. Hu, and R. R. Kompella. 2013. On the impact of packet spraying in data center networks. InIEEE INFOCOM. 2130–2138

    2013

  14. [14]

    Adithya Gangidi, Rui Miao, Shengbao Zheng, Sai Jayesh Bondu, Guilherme Goes, Hany Morsy, Rohit Puri, Mohammad Riftadi, Ashmitha Jeevaraj Shetty, Jingyi Yang, et al. 2024. RDMA over Ethernet for distributed training at Meta scale. InACM SIGCOMM. 57–70

    2024

  15. [15]

    Barak Gerstein, Mark Silberstein, and Isaac Keslassy. 2025. Making congestion control robust to per-packet load balancing in datacenters. https://arxiv.org/abs/2509.07907 arXiv:2509.07907

    arXiv 2025

  16. [16]

    J. Hu, J. Huang, W. Lv, Y. Zhou, J. Wang, and T. He. 2019. CAPS: Coding-based adaptive packet spraying to reduce flow completion time in data center.IEEE/ACM Transactions on Networking27, 6 (2019), 2338–2353

    2019

  17. [17]

    Albert Q Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Emma Bou Hanna, Florian Bressand, et al . 2024. Mixtral of experts.arXiv preprint arXiv:2401.04088(2024)

    Pith/arXiv arXiv 2024

  18. [18]

    Zewen Jin, Shengnan Wang, Jiaan Zhu, Hongrui Zhan, Youhui Bai, Lin Zhang, Zhenyu Ming, and Cheng Li. 2025. BigMac: A Communication- Efficient Mixture-of-Experts Model Structure for Fast Training and Inference. arXiv:2502.16927 [cs.LG] https://arxiv.org/abs/2502.16927

    Pith/arXiv arXiv 2025

  19. [19]

    Y. Le, R. Pan, P. Newman, J. Blendin, A. Kabbani, V. Jain, R. Sivaramu, and F. Matus. 2024. STrack: A reliable multipath transport for ai/ml clusters.arXiv preprint arXiv:2407.15266(2024)

    Pith/arXiv arXiv 2024

  20. [20]

    Yiran Lei, Dongjoo Lee, Liangyu Zhao, Daniar Kurniawan, Chan- myeong Kim, Heetaek block Jeong, Changsu Kim, Hyeonseong Choi, Liangcheng Yu, Arvind Krishnamurthy, Justine Sherry, and Eriko Nurvitadhi. 2026. FAST: An Efficient Scheduler for All-to-All GPU Communication. InProceedings of the 23rd USENIX Symposium on Networked Systems Design and Implementatio...

    2026

  21. [21]

    Xudong Liao, Yijun Sun, Han Tian, Xinchen Wan, Yilun Jin, Zilong Wang, Zhenghang Ren, Xinyang Huang, Wenxue Li, Kin Fai Tse, et al

  22. [22]

    Xinyi Liu, Yujie Wang, Fangcheng Fu, Xupeng Miao, Shenhan Zhu, Xiaonan Nie, and Bin Cui. 2025. NetMoE: Accelerating MoE training through dynamic sample placement. InICML

    2025

  23. [23]

    J. Lu, J. Gao, F. Feng, Z. He, M. Zheng, K. Liu, J. He, B. Liao, S. Xu, K. Sun, et al. 2025. Alibaba Stellar: A new generation RDMA network for cloud AI. InACM SIGCOMM. 453–466

    2025

  24. [24]

    NVIDIA. 2023. NVIDIA InfiniBand Adaptive Routing Technol- ogy—Accelerating HPC and AI Applications. https://www.amax .com/content/files/2023/12/NVIDIA-InfiniBand-Adaptive-Routing- Technology-Insights-Whitepaper.pdf

    2023

  25. [25]

    NVIDIA. 2024. NVIDIA Spectrum-X Network Platform Architecture. https://resources.nvidia.com/en-us-networking-ai/nvidia-spectrum- x

    2024

  26. [26]

    NVIDIA. 2026. Megatron-LM: Training Large Language Models with Model Parallelism. https://github.com/NVIDIA/Megatron-LM. Accessed: July 2026

    2026

  27. [27]

    2026.CUDA C++ Programming Guide (Release 13.3)

    NVIDIA Corporation. 2026.CUDA C++ Programming Guide (Release 13.3). NVIDIA Corporation. Published: June 25, 2026

    2026

  28. [28]

    2026.NVIDIA DOCA Software Framework Docu- mentation

    NVIDIA Corporation. 2026.NVIDIA DOCA Software Framework Docu- mentation. NVIDIA. https://docs.nvidia.com/doca/ Accessed: July 2026

    2026

  29. [29]

    NVIDIA Corporation. 2026. NVIDIA NVLink and NVLink Switch. https://www.nvidia.com/en-eu/data-center/nvlink/. 7 E. Cohenet al

    2026

  30. [30]

    Sundararajan Renganathan and Nick McKeown. 2025. Chronos: Prescheduled circuit switching for LLM training. InProceedings of the 2nd Workshop on Networks for AI Computing. ACM, 89–97

    2025

  31. [31]

    P. Rizk. 2023. Turbocharging Generative AI Workloads with NVIDIA Spectrum-X Networking Platform. https://developer.nvidia.com/blog/ turbocharging-ai-workloads-with-nvidia-spectrum-x-networking- platform/

    2023

  32. [32]

    Chenchen Shou, Guyue Liu, Hao Nie, Huaiyu Meng, Yu Zhou, Yimin Jiang, Wenqing Lv, Yelong Xu, Yuanwei Lu, Zhang Chen, et al. 2025. InfiniteHBD: Building datacenter-scale high-bandwidth domain for LLM with optical circuit switching transceivers. InACM SIGCOMM. 1–23

    2025

  33. [33]

    Ultra Ethernet Consortium. 2025. Ultra Ethernet Specification. https: //ultraethernet.org/

    2025

  34. [34]

    Ultra Ethernet Consortium. 2025. Ultra Ethernet™Specification v1.0. https://ultraethernet.org/wp-content/uploads/sites/20/2025/06/UE- Specification-6.11.25.pdf

    2025

  35. [35]

    Voloshin

    M. Voloshin. 2023.Introduction to Congestion Control for RoCE. Tech. Rep. Broadcom Inc. https://docs.broadcom.com/doc/NCC-WP1XX

    2023

  36. [36]

    John von Neumann. 1953. A certain zero-sum two-person game equiv- alent to the optimal assignment problem. InContributions to the Theory of Games (AM-28), Volume II. Princeton University Press, 5–12

    1953

  37. [37]

    Yen-Chieh Wu, Cheng-Shang Chang, Duan-Shin Lee, and H Jonathan Chao. 2026. Dynamic Hierarchical Birkhoff-von Neumann De- composition for All-to-All GPU Communication.arXiv preprint arXiv:2602.22756(2026)

    arXiv 2026

  38. [38]

    Zhang, D

    J. Zhang, D. Zhang, and K. Huang. 2015. Improving datacenter through- put and robustness with Lazy TCP over packet spraying.Computer Communications62 (2015), 23–33. 8

    2015

  39. [2025]

    InACM SIGCOMM

    MixNet: A Runtime Reconfigurable Optical-Electrical Fabric for Distributed Mixture-of-Experts Training. InACM SIGCOMM. 554–574