Pith. sign in

REVIEW 3 major objections 6 minor 45 references

PRIME: Pseudo-Random Integrated Multi-Part Entropy for Adaptive Packet Spraying in AI/ML Data centers

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read PRIME claims to cut tail flow completion time in AI/ML data-center networks by spraying packets with topology-aware pseudo-random round-robin entropy values and severity-aware congestion history.

desk verdict Plausible incremental design for spray-based load balancing, but the evaluation is unauditable and the manuscript has a missing theorem and non-executable pseudocode. read the letter →

arxiv 2507.23012 v1 pith:QOUJOP2E submitted 2025-07-30 cs.NI

Ashkan Sobhani , Sogand Sadrhaghighi , Xingjun Chu This is my paper · ORCID
classification cs.NI
keywords loadbalancingpacketsprayingAI/MLdatacentersflowcompletiontimemulti-partentropyvaluecongestion-awareroutingFatTreetopologyround-robinscheduling
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

This paper argues that existing packet-spraying load balancers for AI/ML data-center networks build up switch buffers, rely on stale ACK feedback, and handle asymmetric link conditions poorly. It presents PRIME, a sender-driven scheme that sprays packets across all shortest paths using a pseudo-random round-robin over topology-derived multi-part entropy values, and that keeps a severity- and time-aware congestion history to steer packets around hotspots. The claim, supported by packet-level simulations on 2- and 3-tier FatTree topologies, is that PRIME reduces maximum flow completion time by up to 15% for permutation traffic and up to 27% in network-degradation scenarios compared to adaptive routing and REPS. A sympathetic reader would care because tail flow completion time is what limits distributed AI training, so a switch-compatible way to cut queue buildup could translate directly into faster job completion.

What carries the argument

The central object is the multi-part entropy value (MP-EV), a header field divided into as many parts as the topology has tiers; each part encodes the index of the uplink port to use at that tier's switches. It is generated by pseudo-randomized round-robin: per part, an array holds a permutation of port indices and a counter steps through it, reshuffling with the Fisher-Yates algorithm after each wrap-around, while counters for higher tiers advance only when lower-tier counters wrap. Congestion history is the second mechanism: an array indexed by EV stores penalties, set high for NACK (severe congestion) and low for ECN (early congestion), and decayed on every send so paths recover. Together these let a host spray packets uniformly while avoiding currently congested paths, without switch changes beyond ECMP-like EV extraction.

What would settle it

Run a 2-tier FatTree experiment where many flows are all destined to the same leaf switch while PRIME, or its congestion-oblivious variant, spreads packets uniformly over uplinks; if the destination leaf's downlink queue grows deep and maximum flow completion time degrades despite perfectly balanced uplinks, the uplink-balance premise is false.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that a packet-spraying scheme can get better load balance by replacing hash-generated entropy values with a pseudo-randomized round-robin over a multi-part entropy value (MP-EV), where each part names an uplink port at one tier of the Clos topology. Each part cycles through a shuffled permutation of uplink ports, and the permutations are reshuffled after each full cycle to avoid synchronization between flows. On top of this uniform base, PRIME maintains per-path penalties that respond differently to ECN marks and NACK/trimmed-packet signals, with decay so that old congestion information is forgotten. The claimed result is lower buffer occupancy and lower tail flow completion time than REPS and adaptive routing in symmetric, degraded, link-failure, and mixed-traffic settings.

Load-bearing premise

The load-bearing premise is the paper's unproved claim that evenly balancing traffic across all uplinks at every tier is enough to keep downlinks from becoming overloaded, so no per-downlink state is needed.

Editorial extensions

If this is right

  • In symmetric 2-tier and 3-tier FatTree topologies with permutation traffic, PRIME reports up to 15% lower maximum flow completion time than REPS and adaptive routing, with the largest gains for small flows at high bandwidth.
  • Under link degradation, where 25% of links are dropped to 100 Gbps to model in-network-computing traffic, PRIME reports up to 27% improvement over REPS for 4 MB flows.
  • PRIME keeps switch queues shallower than REPS and adaptive routing, and this reduction in buffer occupancy is the mechanism behind its flow-completion-time gains.
  • When ECMP and sprayed traffic coexist under strict-priority or weighted round-robin scheduling, PRIME protects both sprayed and ECMP flows better than REPS does.
  • A congestion-oblivious variant of PRIME matches PRIME in symmetric, uncongested networks, showing that the uniform round-robin spray itself, not the congestion reaction, delivers the baseline benefit.

Reading between the lines

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

  • The paper leaves open whether its uplink-balance premise extends to non-Clos topologies; if it does, the same MP-EV idea could be adapted to dragonfly and other symmetric designs by redefining the mapping from EV parts to links.
  • PRIME's reported resilience to ACK coalescing suggests its feedback path tolerates coarse or delayed congestion signals, so a testable extension is to vary the ACK-coalescing factor and measure whether PRIME's advantage over REPS grows as ACKs become rarer.
  • A direct experimental check would be to implement PRIME in NIC hardware or a programmable switch and compare queue-depth traces against the simulation results, since the reported gains rest entirely on a proprietary simulator.
  • The paper's biggest open question is the missing theorem that balancing uplinks prevents downlink overload; supplying a proof or a counterexample for FatTree would determine whether per-downlink state is truly unnecessary.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes PRIME, a hybrid host-switch packet-spraying load-balancing scheme for AI/ML data-center networks. PRIME generates a multi-part entropy value (MP-EV) whose parts select uplink ports at each network tier, using a pseudo-randomized round-robin procedure with periodic reshuffling, and augments this with a congestion-history mechanism that penalizes paths based on ECN or NACK feedback with different severities and a decay function to avoid stale information. The paper evaluates PRIME in a proprietary packet-level simulator across 2-tier and 3-tier FatTree topologies, comparing against REPS, AR, and a congestion-oblivious variant (CO-PRIME). The central reported claims are up to 15% improvement for permutation traffic and up to 27% improvement in network degradation scenarios, as well as consistent outperformance across the tested settings.

Significance. If the reported results are reproducible and the design is fully specified, PRIME would be a useful contribution to packet-spraying load balancing for AI/ML workloads, particularly for its stated goal of reducing buffer inflation and mitigating the stale-feedback problem under ACK coalescing. The paper does several things right: it evaluates against standard baselines (REPS, AR), considers multiple topologies and bandwidths, separates the pure spraying effect via CO-PRIME, and addresses a practically relevant issue (ACK coalescing). However, the contribution is currently weakened by the absence of a key promised theorem, an inconsistent and non-executable pseudocode specification, and the lack of essential parameter values and reproducibility artifacts. These issues are central because the headline claims are empirical and depend on the exact algorithm and configuration being auditable.

major comments (3)
  1. [§III-D (Multi-part entropy, MP-EV)] The text states that 'balancing the traffic in the uplinks, prevent downlinks to be overloaded in Section ?? (Theorem ??)', but no theorem or proof appears anywhere in the manuscript. This is a load-bearing justification for the MP-EV design: the paper later claims that PRIME 'outperforms alternative packet spraying approaches in all scenarios and under all traffic patterns.' In a FatTree, balanced uplink utilization does not by itself bound downlink queue lengths under skewed or all-to-one traffic, so the conditions under which the claimed property holds must be stated and proved. The authors should either supply a rigorous theorem with explicit assumptions, or temper the generality of the performance claims to the scenarios that are actually supported.
  2. [§III-D and Algorithm 2] Algorithm 2 is not consistent with the prose and with Figure 5. Lines 5–8 of Algorithm 2 select a random index and increment the counter for that random part, whereas the surrounding text and Figure 5 describe a dependent-counter procedure in which the counter for a higher-order part is incremented only when the immediate lower-order counter wraps around. As printed, Algorithm 2 is not executable and does not specify the mechanism that was actually evaluated. The pseudocode must be corrected to match the dependent-counter algorithm, and the shuffle/wraparound behavior must be defined unambiguously.
  3. [§IV (Evaluation setup) and §IV-A (Results)] The evaluation does not report values for the key adaptive parameters: P_ECN, P_NACK, the congestion-decay/drainage schedule, and the ACK-coalescing factor (the latter is stated as 'every 4 packets' in the setup, but the former three are not given anywhere). No seeds, confidence intervals, or sensitivity analyses are provided, and the simulator is proprietary and unreleased. Because the adaptive mechanism can produce a wide range of outcomes depending on these parameters, the headline 'up to 15%' and 'up to 27%' margins are not auditable. The authors should report the exact parameter values used, provide a sensitivity analysis over at least P_ECN, P_NACK, and decay rate, and make the simulator or a reproducible artifact available whenever possible.
minor comments (6)
  1. [Abstract] The phrase 'mix of ordered an unordered traffic' should read 'mix of ordered and unordered traffic'.
  2. [§I Contributions] The last contribution bullet contains an unexplained citation '[7]' at the end of the sentence 'Our results indicate that PRIME outperforms alternative packet spraying approaches in all scenarios and under all traffic patterns [7].' This appears to be a misplaced reference and should be removed or corrected.
  3. [§IV-A (Buffer analysis, Fig. 9)] The text describing Fig. 9 is internally contradictory: it says PRIME 'exhibits significantly lower queue depths' for Fig. 9(a), but then says Fig. 9(b) 'further highlights that REPS maintains a consistently lower queue depth.' The captions and discussion need to be reconciled so the reader knows which scheme has lower buffer occupancy in which subplot.
  4. [§III-D] The unresolved cross-references 'Section ?? (Theorem ??)' appear in the main text and should be fixed or removed; leaving them makes the manuscript look unfinished and masks the missing theorem noted above.
  5. [§IV (Evaluation setup)] The setup text says 'all links have a delay of 600 us'; this is unusually high for a data-center link (typical values are tens of microseconds), and the value should be justified or corrected, as it affects BDP calculations and ECN thresholds.
  6. [§VI Discussion] The discussion of packet reordering says PRIME 'causes minimal re-ordering' and then immediately says 'PRIME results in minimal reordering'; this is redundant and should be condensed. More importantly, the argument that low variance among queue lengths implies minimal reordering is asserted rather than demonstrated, and should be supported with a measurement or a more careful reasoning.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: PRIME's claims are empirical simulator comparisons against external baselines and an ablation; the gaps are auditability and missing-support issues, not definitional reductions.

full rationale

PRIME's central claims are empirical: maximum-FCT ratios measured in a proprietary packet-level simulator, compared against external baselines (REPS, AR) and an internal ablation (CO-PRIME). I found no step where an output quantity is equated with an input by construction, and no fitted parameter is renamed as a prediction. The EV generation and congestion-penalty rules are presented as design decisions; the simulator then measures their consequences. Missing values for P_ECN, P_NACK, the decay schedule, seeds, and confidence intervals make the results hard to audit, and Algorithm 2 is internally inconsistent with the dependent-counter description in Figure 5, but these are reproducibility and correctness concerns, not circularity. The dangling citation '[7]' after 'Our results indicate...' is not a self-citation: reference [7] is SMaRTT-REPS by Bonato et al., disjoint from the current authors. The missing 'Theorem ??' asserting that balancing uplinks prevents downlink overload is an unsupported premise, but an unproved premise is not a derivation that reduces to its own input. Accordingly, the circularity score is 0.

Assumptions & free parameters 6 free parameters · 7 assumptions · 1 invented entities

The central claim rests on a simulation study with several unstated or unreported design parameters (penalties, decay, ECN thresholds, ACK coalescing factor). The most load-bearing assumption is the unproved uplink-to-downlink balancing theorem. The MP-EV header is a new protocol construct with no independent validation.

free parameters (6)
  • PECN (ECN path penalty) = not reported
    Penalty applied to paths whose ACK carries ECN; magnitude and units are not given, and no sensitivity analysis is provided.
  • PNACK (NACK path penalty) = not reported (stated PNACK >> PECN)
    Penalty for paths associated with NACKs; chosen by hand to reflect severe congestion, exact value unreported.
  • Congestion decay rate / drainage rate = not reported (stated close to PECN)
    Decays penalties over time to avoid stale history; the exact function is not specified, affecting path reuse timing.
  • ECN thresholds Kmin, Kmax = 0.25 BDP, 0.75 BDP
    Switch marking thresholds chosen for experiments; results may depend on these values.
  • ACK coalescing factor = 4 packets
    Feedback granularity is set to 4, but no comparison with other coalescing factors is reported.
  • Shuffle random seed = not specified (recommended random per host)
    Controls synchronization between flows; no distribution mechanism is described.
assumptions (7)
  • domain assumption Hash functions cannot produce perfectly uniform output values, causing non-uniform port usage.
    Used in Section III-A to motivate replacing hash-based EV generation with round-robin.
  • standard math Random packet spraying leads to arbitrary-length queue buildup via the balls-into-bins model.
    Invoked in Section II to justify controlled spraying rather than fully random spraying.
  • domain assumption Dequeue ECN marking is supported on most switch hardware and reacts faster than enqueue marking.
    Section III-B assumes switches mark ECN at dequeue, citing [29].
  • ad hoc to paper Balancing uplink traffic at each tier prevents downlink overload.
    Section III-D asserts this with a missing theorem reference (Section ??, Theorem ??); it is load-bearing and unproved.
  • domain assumption ACK coalescing delays and coarsens congestion feedback, causing stale load-balancing decisions.
    The paper's motivation (Section III-A) assumes this, but no experiment isolates the ACK coalescing factor.
  • domain assumption NACK/trimmed packets indicate more severe congestion than ECN-marked ACKs.
    Used to justify PNACK >> PECN in Section III-B; reasonable but a modeling choice without independent validation.
  • domain assumption In FatTree/Clos topologies, a path is fully determined by selecting one uplink port per tier.
    Basis for MP-EV structure in Section III-D; depends on symmetric Clos and shortest-path routing.
invented entities (1)
  • Multi-part entropy value (MP-EV) header field
    purpose: Carries one entropy value per network tier so each switch tier extracts its own uplink port index without per-flow state.
    A new protocol construct; feasibility depends on spare header bits and switch parsing of tier-specific EV lengths, neither validated outside the proprietary simulator.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PRIME: Pseudo-Random Integrated Multi-Part Entropy for Adaptive Packet Spraying in AI/ML Data centers." pith.science (2026). https://pith.science/paper/QOUJOP2E

@misc{pith2026250723012,
  author       = {Pith},
  title        = {Pith review of: PRIME: Pseudo-Random Integrated Multi-Part Entropy for Adaptive Packet Spraying in AI/ML Data centers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QOUJOP2E}},
  note         = {Machine review of arXiv:2507.23012}
}
read the original abstract

Large-scale distributed training in production data centers place significant demands on network infrastructure. In particular, significant load balancing challenges arise when processing AI/ML workloads, consisting of low-entropy, bursty and long-lived flows. Existing solutions designed for Ethernet, such as Equal-Cost Multi-Path (ECMP) struggle to maintain high network utilization. While major industry players (e.g., Ultra Ethernet Consortium) and parts of academia have proposed packet spraying to enhance AI/ML workload performance, we argue that existing packet spraying solutions lead to buffer inflation over time, negatively affecting network performance. Specifically, when ACK coalescing is used, these solutions lead to stale information, degrading network performance. Additionally, in asymmetric network conditions- such as mix of ordered an unordered traffic, or link degradation and failures- existing packet spraying solutions often lead to increased tail latency. In this paper, we present the design and evaluation of PRIME, a pseudo-randomized round-robin approach to packet spraying that considers the network topology to optimize load distribution and performance. PRIME uses congestion as an indicator to re-balance the load. To this extent, PRIME takes into account various congestion signals, accounting for congestion severity, and their decay times to avoid network hotspots. We extensively evaluated PRIME using large-scale production-level simulator. Our results indicate that, compared to existing solutions, PRIME leads to up to 15% improvement for permutation traffic and up to 27% improvement in network degradation scenarios

Figures

Figures reproduced from arXiv: 2507.23012 by the authors.

Figure 1
Figure 1. Example of hash function limitation in EV generation. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Load distribution of 8 flows across different ports of a leaf switch under varying degradation levels when using REPS. the collective completion time, specified by the tail latency. This means that even if most flows are finished quickly, the overall performance is bottle necked by the stragglers. Unfair penalization of flows during asymmetric-network conditions. Unlike traditional networks, where fairness en￾sures … view at source ↗
Figure 3
Figure 3. Example of using pseudo-randomized round-robin for gener [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (7 more)
Figure 5
Figure 5. Figure 5: Example of generating two-part EV. packet needs to be transmitted. We employed a dependent implementation, where the counter for the higher order part (i.e., EV for the higher tier of the network) is incremented once the counter for the immediate lower order is exhaust…
Figure 6
Figure 6. Figure 6: Permutation traffic in a 2-tier FatTree topology with 2048 hosts under varying link bandwidths. (a) 100 Gbps (b) 400 Gbps (c) 800 Gbps [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Permutation traffic in a 3-tier FatTree topology with 1024 hosts under varying link bandwidths. transmission time, resulting in a larger deviation from the ideal FCT. Larger flows, however, are primarily influenced by transmission time, bringing their FCT closer to the…
Figure 8
Figure 8. Figure 8: Average FCT for 8 MB flows in 3-tier FatTree topology with 800 Gbps. Buffer analysis. To further explain why PRIME achieves a lower Max-FCT compared to REPS and AR, [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 10
Figure 10. Figure 10: Link failure. Link degradation. In this experiment, we simulate link degra￾dation, which could occur due to multiple factors such as: SerDes degradation, coexistence of strict-ordered traffic (e.g., In-Network Computing traffic) and unordered traffic, hardware failure…
Figure 12
Figure 12. Figure 12: illustrates the comparison results. The following observations are in order. First, for both algorithms (i.e., REPS and PRIME), ECMP traffic has its lowest maximum FCT, when strict priority is used. The effect on ECMP maximum FCT is due to buffer build up in the netwo…
Figure 11
Figure 11. Figure 11: Link degradation. Mixed flow (coexistence of packet sprayed with non￾sprayed traffic). In these experiments, the performance of PRIME and REPS are evaluated within a 3-tier FatTree topology, under conditions where both ECMP traffic (i.e., non-sprayed) and packet spray…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

45 extracted references · 42 canonical work pages

  1. [7]

    Ethereal: Divide and Conquer Network Load Balancing in Large-Scale Distributed Training

    T. Bonato, A. Kabbani, D. De Sensi, R. Pan, Y . Le, C. Raiciu, M. Han- dley, T. Schneider, N. Black, A. Ghalayini, D. Alves, M. Papamichael, A. Caulfield, and T. Hoefler, “SMaRTT-REPS: Sender-based marked rapidly-adapting trimmed and timed transport with recycled entropies,” arXiv preprint arXiv:2407.00550 , 2024

    work page Pith review arXiv 2024

  2. [34]

    Cognitive routing in the tomahawk 5 data center switch,

    P. D. V . Mohan Kalkunte, Niranjan Vaidya, “Cognitive routing in the tomahawk 5 data center switch,” https://www.broadcom.com/blog/ cognitive-routing-in-the-tomahawk-5-data-center-switch, 2023, [On- line; accessed 31-Mar-2025]

    work page 2023

  3. [2]

    Flexflow: A flexible dataflow accelerator architecture for convolutional neural networks,

    W. Lu, G. Yan, J. Li, S. Gong, Y . Han, and X. Li, “Flexflow: A flexible dataflow accelerator architecture for convolutional neural networks,” in Proc. IEEE International Symposium on High Performance Computer Architecture (HPCA), Feb 2017

    work page 2017

  4. [3]

    Hammingmesh: a network topology for large-scale deep learning,

    T. Hoefler, T. Bonato, D. De Sensi, S. Di Girolamo, S. Li, M. Hed- des, J. Belk, D. Goel, M. Castro, and S. Scott, “Hammingmesh: a network topology for large-scale deep learning,” in Proc. International Conference for High Performance Computing, Networking, Storage and Analysis (SC), Feb 2022

    work page 2022

  5. [4]

    Is network the bottleneck of distributed training?

    Z. Zhang, C. Chang, H. Lin, Y . Wang, R. Arora, and X. Jin, “Is network the bottleneck of distributed training?” in Proc. of the Workshop on Network Meets AI & ML (NetAI) , Aug 2020

    work page 2020

  6. [5]

    CONGA: Distributed congestion-aware load balancing for datacenters,

    M. Alizadeh, T. Edsall, S. Dharmapurikar, R. Vaidyanathan, K. Chu, A. Fingerhut, V . T. Lam, F. Matus, R. Pan, N. Yadav et al., “CONGA: Distributed congestion-aware load balancing for datacenters,” in Proc. in Special Interest Group on Data Communication (SIGCOMM) , Aug 2014

    work page 2014

  7. [6]

    Data center tcp (dctcp),

    M. Alizadeh, A. Greenberg, D. A. Maltz, J. Padhye, P. Patel, B. Prab- hakar, S. Sengupta, and M. Sridharan, “Data center tcp (dctcp),” in Proc. in Special Interest Group on Data Communication (SIGCOMM) , Aug 2010

    work page 2010

  8. [8]

    Datacenter Ethernet and RDMA: Issues at hyperscale,

    T. Hoefler, D. Roweth, K. Underwood, B. Alverson, M. Griswold, V . Tabatabaee, M. Kalkunte, S. Anubolu, S. Shen, A. Kabbani, M. McLaren, and S. Scott, “Datacenter Ethernet and RDMA: Issues at hyperscale,” 2023. [Online]. Available: https://arxiv.org/abs/2302.03337

    arXiv 2023

Show all 45 references
  1. [9]

    Handling inter-DC/Edge AI-related network traffic: Problem statement,

    A. Fressancourt, L. Iannone, Z. Lou, and D. Trossen, “Handling inter-DC/Edge AI-related network traffic: Problem statement,” Internet Engineering Task Force, Internet-Draft draft-aft-ai-traffic-00, oct 2024. [Online]. Available: https://datatracker.ietf.org/doc/draft-aft-ai-tr...

    2024

  2. [10]

    Communication optimization strategies for distributed deep neural network training: A survey,

    S. Ouyang, D. Dong, Y . Xu, and L. Xiao, “Communication optimization strategies for distributed deep neural network training: A survey,” Journal of Parallel and Distributed Computing , vol. 149, 2021

    2021

  3. [11]

    Network load balancing with parallel flowlets for ai training clusters,

    P. Cao, W. Cheng, S. Zhao, and Y . Xiong, “Network load balancing with parallel flowlets for ai training clusters,” in Proc. in SIGCOMM Workshop on Networks for AI Computing , Aug 2024

    2024

  4. [12]

    Multipath issues in unicast and multicast next-hop selection,

    D. Thaler and C. Hopps, “Multipath issues in unicast and multicast next-hop selection,” Tech. Rep., 2000

    2000

  5. [13]

    Let it flow: Resilient asymmetric load balancing with flowlet switching,

    E. Vanini, R. Pan, M. Alizadeh, P. Taheri, and T. Edsall, “Let it flow: Resilient asymmetric load balancing with flowlet switching,” in Proc. USENIX Symposium on Networked Systems Design and Implementation (NSDI), March 2017

    2017

  6. [14]

    Mitigating packet reorder- ing for random packet spraying in data center networks,

    J. Huang, W. Lyu, W. Li, J. Wang, and T. He, “Mitigating packet reorder- ing for random packet spraying in data center networks,” IEEE/ACM Transactions on Networking , vol. 29, no. 3, 2021

    2021

  7. [15]

    Hedera: Dynamic flow scheduling for data center networks,

    M. Al-Fares, S. Radhakrishnan, B. Raghavan, N. Huang, and A. Vahdat, “Hedera: Dynamic flow scheduling for data center networks,” in proc. USENIX Symposium on Networked Systems Design and Implementation (NSDI), Apr 2010

    2010

  8. [16]

    The new era needs a new network,

    U. E. Consortium, “The new era needs a new network,” https:// ultraethernet.org/, 2008, [Online; accessed 18-Feb-2025]

    2008

  9. [17]

    Empowering ai workloads in ultra ethernet consortium,

    J. Metz, “Empowering ai workloads in ultra ethernet consortium,” in Proc. IEEE Photonics Society Summer Topicals Meeting Series (SUM) , July 2024

    2024

  10. [18]

    On the impact of packet spraying in data center networks,

    A. Dixit, P. Prakash, Y . C. Hu, and R. R. Kompella, “On the impact of packet spraying in data center networks,” in Proc. IEEE International Conference on Computer Communications (INFOCOM , Apr 2013

    2013

  11. [19]

    Drill: Micro load balancing for low-latency data center networks,

    S. Ghorbani, Z. Yang, P. B. Godfrey, Y . Ganjali, and A. Firoozshahian, “Drill: Micro load balancing for low-latency data center networks,” in Proc. ACM Special Interest Group on Data Communication (SIG- COMM), Aug 2017

    2017

  12. [20]

    STrack:A reliable multipath transport for ai/ml clusters,

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

    2024 arXiv

  13. [21]

    Random early detection gateways for congestion avoidance,

    S. Floyd and V . Jacobson, “Random early detection gateways for congestion avoidance,” IEEE/ACM Transactions on networking , vol. 1, no. 4, 1993

    1993

  14. [22]

    Re-architecting datacenter networks and stacks for low latency and high performance,

    M. Handley, C. Raiciu, A. Agache, A. V oinescu, A. W. Moore, G. An- tichi, and M. Wójcik, “Re-architecting datacenter networks and stacks for low latency and high performance,” in Proc. ACM Special Interest Group on Data Communication (SIGCOMM) , Aug 2017

    2017

  15. [23]

    Multi-Path transport for RDMA in datacenters,

    Y . Lu, G. Chen, B. Li, K. Tan, Y . Xiong, P. Cheng, J. Zhang, E. Chen, and T. Moscibroda, “Multi-Path transport for RDMA in datacenters,” in Proc. USENIX symposium on networked systems design and implemen- tation (NSDI), Apr 2018

    2018

  16. [24]

    PLB: Congestion signals are simple and effective for network load balancing,

    M. A. Qureshi, Y . Cheng, Q. Yin, Q. Fu, G. Kumar, M. Moshref, J. Yan, V . Jacobson, D. Wetherall, and A. Kabbani, “PLB: Congestion signals are simple and effective for network load balancing,” in Proc. ACM Special Interest Group on Data Communication (SIGCOMM), Aug 2022

    2022

  17. [25]

    Best nexthop load balancing algorithm with inband network telemetry,

    J. Lim, S. Nam, J.-H. Yoo, and J. W.-K. Hong, “Best nexthop load balancing algorithm with inband network telemetry,” in Proc. IEEE International Conference on Network and Service Management (CNSM), Nov 2020

    2020

  18. [26]

    REPS: Recycled entropy packet spraying for adaptive load balancing and failure mitigation,

    T. Bonato, A. Kabbani, A. Ghalayini, M. Papamichael, M. Dohadwala, L. Gianinazzi, M. Khalilov, E. Achermann, D. De Sensi, and T. Hoefler, “REPS: Recycled entropy packet spraying for adaptive load balancing and failure mitigation,” arXiv preprint arXiv:2407.21625v3 , 2025

    2025

  19. [27]

    Per-packet load-balanced, low-latency routing for clos-based data center networks,

    J. Cao, R. Xia, P. Yang, C. Guo, G. Lu, L. Yuan, Y . Zheng, H. Wu, Y . Xiong, and D. Maltz, “Per-packet load-balanced, low-latency routing for clos-based data center networks,” in Proc. ACM conference on Emerging networking experiments and technologies (CoNEXT) , Dec 2013

    2013

  20. [28]

    The addition of explicit congestion notification (ECN) to IP,

    K. Ramakrishnan, S. Floyd, and D. Black, “The addition of explicit congestion notification (ECN) to IP,” Tech. Rep., 2001

    2001

  21. [29]

    Tuning ECN for data center networks,

    H. Wu, J. Ju, G. Lu, C. Guo, Y . Xiong, and Y . Zhang, “Tuning ECN for data center networks,” in Proc. of the international conference on Emerging networking experiments and technologies (CoNECT) , Dec 2012

    2012

  22. [30]

    Fisher–yates shuffle,

    M. Eberl, “Fisher–yates shuffle,” https://ultraethernet.org/, 2016, [On- line; accessed 28-Feb-2025]

    2016

  23. [31]

    SeqBalance: Congestion-aware load balancing with no reordering for roce,

    H. Luo, J. Zhang, M. Yu, Y . Pan, T. Pan, and T. Huang, “SeqBalance: Congestion-aware load balancing with no reordering for roce,” arXiv preprint arXiv:2407.09808, 2024

    2024 arXiv

  24. [32]

    What is bfd?

    S. Hongye, “What is bfd?” https://info.support.huawei.com/info-finder/ encyclopedia/en/BFD.html, 2024, [Online; accessed 31-Mar-2025]

    2024

  25. [33]

    What is dpfr?

    Y . Xiaoli, “What is dpfr?” https://info.support.huawei.com/info-finder/ encyclopedia/en/DPFR.html, 2024, [Online; accessed 31-Mar-2025]

    2024

  26. [35]

    MicroTE: Fine grained traffic engineering for data centers,

    T. Benson, A. Anand, A. Akella, and M. Zhang, “MicroTE: Fine grained traffic engineering for data centers,” in Proc. of the seventh conference on emerging networking experiments and technologies(CoNEXT) , Dec 2011

    2011

  27. [36]

    Mahout: Low-overhead datacenter traffic management using end-host-based elephant detection,

    A. R. Curtis, W. Kim, and P. Yalagandula, “Mahout: Low-overhead datacenter traffic management using end-host-based elephant detection,” in Proc. IEEE International Conference on Computer Communications (INFOCOM), Apr 2011

    2011

  28. [37]

    Planck: Millisecond-scale monitoring and control for commodity networks,

    J. Rasley, B. Stephens, C. Dixon, E. Rozner, W. Felter, K. Agarwal, J. Carter, and R. Fonseca, “Planck: Millisecond-scale monitoring and control for commodity networks,” ACM SIGCOMM Computer Commu- nication Review, vol. 44, no. 4, 2014

    2014

  29. [38]

    Flowbender: Flow- level adaptive routing for improved latency and throughput in datacenter networks,

    A. Kabbani, B. Vamanan, J. Hasan, and F. Duchene, “Flowbender: Flow- level adaptive routing for improved latency and throughput in datacenter networks,” in Proc. ACM International on Conference on emerging Networking Experiments and Technologies (CoNEXT) , Dec 2014

    2014

  30. [39]

    Hula: Scalable load balancing using programmable data planes,

    N. Katta, M. Hira, C. Kim, A. Sivaraman, and J. Rexford, “Hula: Scalable load balancing using programmable data planes,” in Proc. Symposium on SDN Research (SOSR) , Mar 2016

    2016

  31. [40]

    CLOVE: How i learned to stop worrying about the core and love the edge,

    N. Katta, M. Hira, A. Ghag, C. Kim, I. Keslassy, and J. Rexford, “CLOVE: How i learned to stop worrying about the core and love the edge,” in Proc. ACM Workshop on Hot Topics in Networks( HotN ets), Nov 2016

    2016

  32. [41]

    Presto: Edge-based load balancing for fast datacenter networks,

    K. He, E. Rozner, K. Agarwal, W. Felter, J. Carter, and A. Akella, “Presto: Edge-based load balancing for fast datacenter networks,” ACM SIGCOMM Computer Communication Review , vol. 45, no. 4, 2015

    2015

  33. [42]

    Fastpass: A centralized

    J. Perry, A. Ousterhout, H. Balakrishnan, D. Shah, and H. Fugal, “Fastpass: A centralized" zero-queue" datacenter network,” in Proc. ACM Special Interest Group on Data Communication SIGCOMM , Aug 2014

    2014

  34. [43]

    QDAPS: Queueing delay aware packet spraying for load balancing in data center,

    J. Huang, W. Lv, W. Li, J. Wang, and T. He, “QDAPS: Queueing delay aware packet spraying for load balancing in data center,” in Proc. International Conference on Network Protocols (ICNP) . IEEE, Sep 2018

    2018

  35. [44]

    AG: Adaptive switching granular- ity for load balancing with asymmetric topology in data center network,

    J. Liu, J. Huang, W. Li, and J. Wang, “AG: Adaptive switching granular- ity for load balancing with asymmetric topology in data center network,” in Proc. International Conference on Network Protocols (ICNP) , Oct 2019

    2019

  36. [45]

    Congestion signaling (CSIG),

    N. M. J. K. A. Ravi, N. Dukkipati, “Congestion signaling (CSIG),” https://www.ietf.org/archive/id/draft-ravi-ippm-csig-01.html, 2024, [On- line; accessed 31-Mar-2025]

    2024

  37. [46]

    In-band network telemetry (INT) dataplane specification,

    T. P. A. W. Group, “In-band network telemetry (INT) dataplane specification,” https://p4.org/p4-spec/docs/INT_v2_1.pdf, 2020, [Online; accessed 31-Mar-2025]

    2020

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.