pith. sign in

arxiv: 2606.27697 · v1 · pith:23SAADWRnew · submitted 2026-06-26 · 💻 cs.NI

Host-Driven Flowlet Balancing with Segment Routing over IPv6

Ryo Nakamura , Hiroki Kano , Tomoko Okuzawa This is my paper

Pith reviewed 2026-06-29 02:54 UTC · model grok-4.3

classification 💻 cs.NI
keywords flowlet balancingSRv6host-driven load balancingdata center networkstail latencysegment routingECMPin-flight byte estimation
0
0 comments X

The pith

Hosts steer flowlets via SRv6 using an in-flight byte model to cut tail latency 33 percent versus ECMP.

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

The paper proposes a host-driven flowlet balancing method that uses Segment Routing over IPv6 to let end hosts detect flowlets in outgoing traffic and steer them onto specific paths. Switches remain stateless and simply forward based on the SRv6 segments. Each host applies a simple model that estimates in-flight bytes on each available path to distribute flowlets as evenly as possible. Evaluation on a Linux testbed with an SRv6 router shows that, for fixed-size flows, this approach reduces tail latency by 15 percent relative to random flowlet balancing and by 33 percent relative to ECMP. The same method combined with dynamic flowlet timeouts also improves results under two application workloads.

Core claim

A fully host-driven flowlet balancing method with SRv6 allows hosts to detect flowlets and steer them using a simple in-flight byte estimation model, keeping switches stateless; under fixed-size flows this yields 15 percent lower tail latency than random flowlet balancing and 33 percent lower than ECMP, with further gains when paired with dynamic timeouts for real workloads.

What carries the argument

The simple model that estimates in-flight bytes on each path to decide flowlet-to-path assignments, executed at the host and encoded into SRv6 segments.

If this is right

  • Under fixed-size flows the method lowers tail latency 15 percent versus random flowlet balancing and 33 percent versus ECMP.
  • Combining the byte-estimation model with dynamically adjusted flowlet timeouts improves performance on application workloads.
  • Switches operate statelessly as ordinary SRv6 nodes with no per-flow state required.
  • All flowlet detection and path selection logic resides at the sending hosts.

Where Pith is reading between the lines

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

  • The same host-side byte model could be tested with other source-routing mechanisms beyond SRv6 if they allow per-packet path selection.
  • If the byte estimator remains accurate at scale, the approach might reduce the need for in-network flowlet state in larger data-center fabrics.
  • A direct comparison against per-flowlet state kept at the first-hop switch would isolate how much of the gain comes from host visibility versus the estimation model itself.

Load-bearing premise

The simple model that estimates in-flight bytes on each path is sufficiently accurate to guide effective flowlet distribution decisions across paths.

What would settle it

A measurement showing that actual bytes in flight on each path deviate substantially from the host's estimates, producing latency no better than or worse than ECMP.

Figures

Figures reproduced from arXiv: 2606.27697 by Hiroki Kano, Ryo Nakamura, Tomoko Okuzawa.

Figure 1
Figure 1. Figure 1: Flowlet balancing with SRv6. The host detects flowlets [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: A model for estimating in-flight bytes on a path as a [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: depicts our testbed. The experimental topology is shown in Figure 3a; there are four leaf switches, and each is connected to eight hosts and four spine switches with 1 Gbps links. Namely, it is a 2:1 over-subscription. This topology was realized on the physical setup shown in Figure 3b. The router was MX10003 from Juniper Networks that supports Leaf 1 Spine 1 Spine 2 S1 S8 Leaf 2 S9 S16 ... 1Gbps Leaf 3 Sp… view at source ↗
Figure 4
Figure 4. Figure 4: CDF of FCTs with 100KB flows. 2 3 4 5 6 7 flow completion time (msec) 90 95 100 percentile ECMP RPS LetFlow P2C Halflife-RND Halflife-P2C [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Tail FCTs with 100KB flows. 3.01 ms, 28% and 43% reductions compared with LetFlow and ECMP, respectively. These results indicate that the power-of￾two choices algorithm with in-flight byte estimation achieves better balancing than the random path selection. In addition, dynamic FTV adjustment improves the performance compared with the fixed FTV, and RPS results in longer tail latency. C. Results with Reali… view at source ↗
Figure 8
Figure 8. Figure 8: Median FCTs for the Web Search workload. [PITH_FULL_IMAGE:figures/full_fig_p005_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: 99th-percentile FCTs for the Web Search workload. [PITH_FULL_IMAGE:figures/full_fig_p006_9.png] view at source ↗
read the original abstract

This paper proposes a fully host-driven method for flowlet balancing with Segment Routing over IPv6 (SRv6). In modern data center networks, load balancing plays a pivotal role in efficiently utilizing multiple paths. Flowlet balancing offers finer granularity in traffic splitting than ECMP and is therefore expected to achieve higher performance. However, deploying flowlet balancing in practice is still challenging due to the scalability issue of switches having to maintain per-flow state. In our approach, hosts detect flowlets in their outgoing traffic and steer them onto specific paths using SRv6. The switches behave as SRv6 nodes in a stateless manner. Each host distributes its flowlets as evenly as possible across paths. As a metric for this load balancing, we introduce a simple model that estimates in-flight bytes on each path. We implemented the proposed method on Linux and evaluated it on a testbed with an SRv6-capable router. The results show that, under fixed-size flows, the proposed method reduces tail latency by 15% and 33% compared with random flowlet balancing and ECMP, respectively. Furthermore, combining the method with dynamically adjusted flowlet timeouts also improves performance under two application 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

2 major / 1 minor

Summary. The paper proposes a host-driven flowlet balancing scheme for SRv6 networks in which end hosts detect flowlets in outgoing traffic and assign them to paths using a simple model that estimates in-flight bytes per path; switches remain stateless SRv6 nodes. The method is implemented on Linux and evaluated on a testbed, claiming 15% and 33% tail-latency reductions versus random flowlet balancing and ECMP under fixed-size flows, plus further gains when combined with dynamic flowlet timeouts on application workloads.

Significance. If the in-flight estimation model is shown to be accurate and robust, the approach would be significant for enabling fine-grained, scalable load balancing entirely at the host without switch state or hardware changes, addressing a practical deployment barrier for flowlet-based schemes in data-center networks.

major comments (2)
  1. [Abstract / evaluation section] Abstract and evaluation description: the headline claims of 15% and 33% tail-latency reduction rest entirely on hosts using the in-flight-bytes model to make assignment decisions, yet the manuscript provides neither the model's equations or pseudocode, nor any measurement of its estimation error (e.g., against ground-truth in-flight bytes, sensitivity to RTT variation, or cross-traffic). This is load-bearing for the central performance claim.
  2. [Evaluation] Evaluation methodology: the abstract states performance numbers but supplies no information on the number of runs, variance, confidence intervals, or how the testbed traffic was generated and measured. Without these details the reported gains versus random flowlet balancing cannot be assessed for statistical significance or reproducibility.
minor comments (1)
  1. [Abstract] The abstract refers to 'a simple model' without naming the section where the model is defined; adding an explicit forward reference would improve readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive comments. We address each major point below and will revise the manuscript to improve the description of the in-flight-bytes model and the reporting of experimental methodology.

read point-by-point responses

  1. Referee: [Abstract / evaluation section] Abstract and evaluation description: the headline claims of 15% and 33% tail-latency reduction rest entirely on hosts using the in-flight-bytes model to make assignment decisions, yet the manuscript provides neither the model's equations or pseudocode, nor any measurement of its estimation error (e.g., against ground-truth in-flight bytes, sensitivity to RTT variation, or cross-traffic). This is load-bearing for the central performance claim.

    Authors: We agree that the in-flight-bytes model is central to the performance claims and that its details should be explicit. The manuscript describes the model at a high level in Section 3, but we will add the precise equations, pseudocode for path assignment, and new measurements of estimation accuracy (including error against ground truth, sensitivity to RTT, and cross-traffic effects) in a revised evaluation section or dedicated subsection. revision: yes

  2. Referee: [Evaluation] Evaluation methodology: the abstract states performance numbers but supplies no information on the number of runs, variance, confidence intervals, or how the testbed traffic was generated and measured. Without these details the reported gains versus random flowlet balancing cannot be assessed for statistical significance or reproducibility.

    Authors: We concur that these methodological details are required for assessing statistical significance and reproducibility. In the revised manuscript we will report the number of runs performed, observed variance, confidence intervals on the latency reductions, and a complete description of traffic generation and measurement procedures in the testbed. revision: yes

Circularity Check

0 steps flagged

No circularity: claims rest on testbed measurements with no equations or fitted derivations

full rationale

The paper introduces a simple model for estimating in-flight bytes but provides no equations, derivations, or parameter-fitting steps in the supplied text. All reported performance gains (15%/33% tail-latency reductions) are presented as direct outcomes of testbed experiments rather than any mathematical reduction or self-referential prediction. No self-citations, ansatzes, or uniqueness theorems are invoked to justify core results. The derivation chain is therefore empty; the work is self-contained against external benchmarks and receives the default non-finding.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central performance claims rest on the accuracy of an unspecified simple model for in-flight bytes and on the assumption that hosts can reliably detect flowlets in outgoing traffic. No free parameters or invented entities are described in the abstract.

axioms (1)
  • domain assumption A simple model can estimate in-flight bytes on each path with enough fidelity to support load-balancing decisions.
    Invoked to justify the balancing metric; details absent from abstract.

pith-pipeline@v0.9.1-grok · 5740 in / 1139 out tokens · 50523 ms · 2026-06-29T02:54:59.636664+00:00 · methodology

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

29 extracted references

  1. [1]

    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,” inProceedings of the ACM SIGCOMM 2022 Conference, ser. SIGCOMM ’22, 2022, p. 207–218

    2022

  2. [2]

    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,” in2013 Proceedings IEEE INFOCOM, 2013, pp. 2130–2138

    2013

  3. [3]

    Wcmp: weighted cost multipathing for improved fairness in data centers,

    J. Zhou, M. Tewari, M. Zhu, A. Kabbani, L. Poutievski, A. Singh, and A. Vahdat, “Wcmp: weighted cost multipathing for improved fairness in data centers,” inProceedings of the Ninth European Conference on Computer Systems, ser. EuroSys ’14, 2014

    2014

  4. [4]

    Dynamic load balancing without packet reordering,

    S. Kandula, D. Katabi, S. Sinha, and A. Berger, “Dynamic load balancing without packet reordering,”SIGCOMM Comput. Commun. Rev., vol. 37, no. 2, p. 51–62, Mar. 2007

    2007

  5. [5]

    Configure Flowset Table Size in DLB Flowlet Mode — Junos OS Evolved — Juniper Networks,

    “Configure Flowset Table Size in DLB Flowlet Mode — Junos OS Evolved — Juniper Networks,” Apr. 2024. [Online]. Available: https://www.juniper.net/documentation/us/en/software/junos/ai-ml- evo/interfaces-ethernet-switches/topics/topic-map/configure-flowset- table-size.html

    2024

  6. [6]

    Alibaba hpn: A data center network for large language model training,

    K. Qian, Y . Xi, J. Cao, J. Gao, Y . Xu, Y . Guan, B. Fu, X. Shi, F. Zhu, R. Miao, C. Wang, P. Wang, P. Zhang, X. Zeng, E. Ruan, Z. Yao, E. Zhai, and D. Cai, “Alibaba hpn: A data center network for large language model training,” inProceedings of the ACM SIGCOMM 2024 Conference, ser. ACM SIGCOMM ’24, 2024, p. 691–706

    2024

  7. [7]

    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,” in Proceedings of the 2015 ACM Conference on Special Interest Group on Data Communication, ser. SIGCOMM ’15, 2015, p. 465–478

    2015

  8. [8]

    Clove: Congestion-aware load balancing at the virtual edge,

    N. Katta, A. Ghag, M. Hira, I. Keslassy, A. Bergman, C. Kim, and J. Rexford, “Clove: Congestion-aware load balancing at the virtual edge,” inProceedings of the 13th International Conference on Emerging Networking EXperiments and Technologies, ser. CoNEXT ’17, 2017, p. 323–335

    2017

  9. [9]

    Adaptive network load balancing at the end host for traffic bursts in data centers,

    Q. Shi, H. Huang, X. Li, C. Li, W. Cao, and L. Liu, “Adaptive network load balancing at the end host for traffic bursts in data centers,” in2024 IEEE International Conference on High Performance Computing and Communications (HPCC), Dec 2024, pp. 416–423

    2024

  10. [10]

    Segment Routing over IPv6 (SRv6) Network Programming,

    C. Filsfils, P. Camarillo, J. Leddy, D. V oyer, S. Matsushima, and Z. Li, “Segment Routing over IPv6 (SRv6) Network Programming,” RFC 8986, Feb. 2021

    2021

  11. [11]

    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, and G. Varghese, “Conga: distributed congestion-aware load balancing for datacenters,” inProceedings of the 2014 ACM Conference on SIGCOMM, ser. SIGCOMM ’14, 2014, p. 503–514

    2014

  12. [12]

    Hula: Scal- able load balancing using programmable data planes,

    N. Katta, M. Hira, C. Kim, A. Sivaraman, and J. Rexford, “Hula: Scal- able load balancing using programmable data planes,” inProceedings of the Symposium on SDN Research, ser. SOSR ’16, 2016

    2016

  13. [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,” in14th USENIX Symposium on Networked Systems Design and Implementation (NSDI 17), Boston, MA, Mar. 2017, pp. 407–420

    2017

  14. [14]

    Flex: A flowlet-level load balancing based on load-adaptive timeout in dcn,

    X. Diao, H. Gu, X. Yu, L. Qin, and C. Luo, “Flex: A flowlet-level load balancing based on load-adaptive timeout in dcn,”Future Gener. Comput. Syst., vol. 130, no. C, p. 219–230, May 2022

    2022

  15. [15]

    Halflife: An adaptive flowlet-based load balancer with fading timeout in data center networks,

    S. Liu, Y . Gao, Z. Chen, J. Ye, H. Xu, F. Liang, W. Yan, Z. Tian, Q. Sun, Z. Guo, and Y . Xu, “Halflife: An adaptive flowlet-based load balancer with fading timeout in data center networks,” inProceedings of the Nineteenth European Conference on Computer Systems, ser. EuroSys ’24, 2024, p. 66–81

    2024

  16. [16]

    Cognitive routing in the tomahawk 5 data center switch,

    M. Kalkunte, N. Vaidya, and P. Del Vecchio, “Cognitive routing in the tomahawk 5 data center switch,” Feb. 2023. [Online]. Available: https://www.broadcom.com/blog/cognitive-routing- in-the-tomahawk-5-data-center-switch

    2023

  17. [17]

    Hf2t: Host-based flowlet fine-tuning for rdma load balancing,

    C. Chen, J. Ye, Y . Gao, S. Liu, and Y . Xu, “Hf2t: Host-based flowlet fine-tuning for rdma load balancing,” inProceedings of the 8th Asia- Pacific Workshop on Networking, ser. APNet ’24, 2024, p. 9–15

    2024

  18. [18]

    Ro- celet: Host-based flowlet load balancing for roce,

    H. Luo, J. Zhang, M. Yu, J. Jiang, Y . Pan, T. Pan, and T. Huang, “Ro- celet: Host-based flowlet load balancing for roce,”IEEE Transactions on Networking, vol. 33, no. 4, pp. 1676–1688, 2025

    2025

  19. [19]

    Explicit path control in commodity data centers: Design and applications,

    S. Hu, K. Chen, H. Wu, W. Bai, C. Lan, H. Wang, H. Zhao, and C. Guo, “Explicit path control in commodity data centers: Design and applications,” in12th USENIX Symposium on Networked Systems Design and Implementation (NSDI 15), Oakland, CA, May 2015, pp. 15–28

    2015

  20. [20]

    Sonic for ai with srv6,

    R. Hui, “Sonic for ai with srv6,” Mar. 2025. [Online]. Avail- able: https://www.segment-routing.net/conferences/Paris25-Microsoft- Rita-Hui/

    2025

  21. [21]

    IS-IS Extensions to Support Segment Routing over the IPv6 Data Plane,

    P. Psenak, C. Filsfils, A. Bashandy, B. Decraene, and Z. Hu, “IS-IS Extensions to Support Segment Routing over the IPv6 Data Plane,” RFC 9352, Feb. 2023

    2023

  22. [22]

    OSPFv3 Extensions for Segment Routing over IPv6 (SRv6),

    Z. Li, Z. Hu, K. Talaulikar, and P. Psenak, “OSPFv3 Extensions for Segment Routing over IPv6 (SRv6),” RFC 9513, Dec. 2023

    2023

  23. [23]

    Compressed SRv6 Segment List Encoding,

    W. Cheng, C. Filsfils, Z. Li, B. Decraene, and F. Clad, “Compressed SRv6 Segment List Encoding,” RFC 9800, Jun. 2025

    2025

  24. [24]

    The power of two choices in randomized load balancing,

    M. Mitzenmacher, “The power of two choices in randomized load balancing,”IEEE Transactions on Parallel and Distributed Systems, vol. 12, no. 10, pp. 1094–1104, 2001

    2001

  25. [25]

    On getting tc classifier fully programmable with cls bpf,

    D. Borkmann, “On getting tc classifier fully programmable with cls bpf,” inThe Technical Conference on Linux Networking (Netdev 1.1), Nov. 2016. [Online]. Available: https://netdevconf.info/1.1/talk-getting- tc-classifier-fully-programmable-clsbpf-daniel-borkmann.html

    2016

  26. [26]

    FRRouting,

    “FRRouting,” 2025. [Online]. Available: https://frrouting.org/

    2025

  27. [27]

    upa/flowperf: flowperf: A flow-based network benchmark tool,

    “upa/flowperf: flowperf: A flow-based network benchmark tool,” 2025. [Online]. Available: https://github.com/upa/flowperf

    2025

  28. [28]

    Inside the social network’s (datacenter) network,

    A. Roy, H. Zeng, J. Bagga, G. Porter, and A. C. Snoeren, “Inside the social network’s (datacenter) network,” inProceedings of the 2015 ACM Conference on Special Interest Group on Data Communication, ser. SIGCOMM ’15, 2015, p. 123–137

    2015

  29. [29]

    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 Proceedings of the ACM SIGCOMM 2010 Conference, ser. SIGCOMM ’10, 2010, p. 63–74

    2010