pith. sign in

arxiv: 2606.12963 · v1 · pith:PKR4IWT7new · submitted 2026-06-11 · 💻 cs.NI · cs.DC· cs.ET

ScaleAcross: Designing Multi-Data-Center Infrastructure for Geo-Distributed AI Training

Naved Inam , Aryan Alpesh Bhavsar , Masabattula Teja Nikhil , Sidharth Sharma This is my paper

Pith reviewed 2026-06-27 05:39 UTC · model grok-4.3

classification 💻 cs.NI cs.DCcs.ET
keywords geo-distributed AI trainingEVPN-VXLANmulti-data-center infrastructureAllReduceParameter ServerWAN emulationContainerLabFRRouting
0
0 comments X

The pith

An EVPN-VXLAN emulation framework allows study of geo-distributed AI training workloads over wide-area networks with commodity tools.

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

The paper develops an emulation framework for connecting multiple data centers to support AI model training across geographic sites. It layers VXLAN overlays on EVPN for inter-site links, built on ContainerLab and FRRouting, and adds ECMP routing, BFD detection, and queue-pair-aware traffic handling. These elements let researchers run AllReduce and Parameter Server patterns under emulated wide-area conditions to observe synchronization behavior. The approach stays compatible with standard hardware while addressing latency and data-sovereignty constraints. If the emulation holds, it offers a practical way to test and tune infrastructure for large-scale distributed training without physical multi-site builds.

Core claim

The authors present a framework that combines VXLAN overlays with EVPN-based inter-data-center connectivity, implemented using ContainerLab and FRRouting. It incorporates Equal-Cost Multi-Path routing, Bidirectional Forwarding Detection, and a queue-pair-aware traffic distribution mechanism to handle synchronization-intensive workloads such as AllReduce and Parameter Server communication under realistic wide-area emulation.

What carries the argument

EVPN-VXLAN overlay network with ECMP routing, BFD, and queue-pair-aware traffic distribution for multi-data-center AI workload connectivity.

If this is right

  • Traffic distribution improves for synchronization-intensive AI workloads while using standard routing features.
  • BFD provides faster failure detection across data-center boundaries.
  • The setup remains compatible with commodity infrastructure without custom hardware.
  • Reproducible emulation yields insights into traffic and resilience behavior for geo-distributed training.

Where Pith is reading between the lines

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

  • The emulation approach could reduce experimental costs by substituting for physical multi-site deployments in early-stage infrastructure design.
  • Insights on queue-pair-aware distribution might extend to optimizing other collective communication patterns in distributed systems.
  • Integration with real AI training frameworks could measure end-to-end effects on convergence time under emulated WAN conditions.

Load-bearing premise

The ContainerLab plus FRRouting emulation accurately reproduces latency, jitter, and packet-loss behavior of real wide-area networks when running AllReduce and Parameter Server workloads at scale.

What would settle it

Direct comparison of the same AllReduce and Parameter Server workloads on a physical multi-data-center testbed showing substantially different traffic distribution, latency, or packet-loss patterns than observed in the emulation.

Figures

Figures reproduced from arXiv: 2606.12963 by Aryan Alpesh Bhavsar, Masabattula Teja Nikhil, Naved Inam, Sidharth Sharma.

Figure 1
Figure 1. Figure 1: Experimental topology showing simultaneous deployment of Parameter Server (M1) and All Reduce (M2) architectures across the emulated geo-distributed data centers. Although statistically effective, previous work [8] identified production scenarios in which different QPs commu￾nicating between the same source-destination GPU pair receive identical source ports. This produces identical packet 5-tuples, causin… view at source ↗
Figure 2
Figure 2. Figure 2: Example Containerlab YAML topology definition [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Containerlab deployment output confirming successful startup of all nodes and Debian-based hosts for traffic generation and distributed training. Links were defined to emulate both intra￾datacenter and inter-datacenter connectivity [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Container-based Soft-RoCE VM implementation [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: VXLAN+EVPN control/data plane operations: model training using Parameter Server (PS) architecture in geo-distributed AI data centers. 4.2.5. Control-plane and data-plane operations [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Prometheus configuration file with SNMP and Ping Exporter targets 5.2. Equal-Cost Multi-Path Routing (ECMP) To evaluate the ECMP routing, multiple traffic flows were generated from d1h1 to d2h2. As shown in the topology diagram ( [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Prometheus showing active monitoring targets [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Ping RTT from Datacenter1 Host1 to Datacenter2 Host1 after introducing artificial delay and jitter using Containerlab’s netem tool [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Ping RTT over time with BFD enabled (10 ms interval, 3 retries). Recovery is achieved in approximately 110 ms [PITH_FULL_IMAGE:figures/full_fig_p013_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: [Left]: At the leaf node (Leaf1, DC1), traffic is evenly distributed across both uplinks to the spine layer. [Right]: At the spine node (Spine1, DC1), traffic is further balanced across two WAN links toward DC2, confirming ECMP functionality at both the aggregation and core layers. 5.3. Link Failure Recovery To evaluate the network’s resilience and convergence time during failures, we conducted a link fai… view at source ↗
Figure 11
Figure 11. Figure 11: Load Factor at Leaf Switch [PITH_FULL_IMAGE:figures/full_fig_p014_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Load Factor at Spine Switch physical topology. The experimental setup, based on the spine-leaf topology described in [PITH_FULL_IMAGE:figures/full_fig_p014_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Ping RTT over time during link failure recovery using default BGP timers. Recovery takes approximately 180 seconds due to slow failure detection [PITH_FULL_IMAGE:figures/full_fig_p015_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Batch-wise gradient computation and synchronization time for Allreduce and Parameter Server architecture [PITH_FULL_IMAGE:figures/full_fig_p016_14.png] view at source ↗
read the original abstract

The rapid growth of AI models and increasing data sovereignty requirements are driving the transition toward geo-distributed AI training across multiple data centers. Such deployments introduce system-level challenges arising from synchronization-intensive communication, cross-site data exchange, and wide-area latency constraints. This paper investigates EVPN--VXLAN as an infrastructure foundation for geo-distributed AI training environments and presents a scalable emulation framework for systematically studying distributed AI workloads under realistic wide-area conditions. The proposed framework combines VXLAN overlays with EVPN-based inter-data-center connectivity and is implemented using ContainerLab and FRRouting (FRR). The framework further incorporates Equal-Cost Multi-Path (ECMP) routing, Bidirectional Forwarding Detection (BFD), and a queue-pair-aware traffic distribution mechanism designed to improve communication behavior for synchronization-intensive AI workloads while preserving compatibility with commodity infrastructure. Using realistic WAN emulation, we characterize communication and system behavior under distributed training workloads employing AllReduce and Parameter Server communication patterns. Results provide insights into traffic distribution, resilience, and infrastructure behavior in geo-distributed AI environments, highlighting the potential of reproducible multi-data-center infrastructure frameworks for scalable distributed AI training.

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 / 2 minor

Summary. The manuscript presents ScaleAcross, a framework for multi-data-center infrastructure supporting geo-distributed AI training. It combines VXLAN overlays with EVPN-based inter-DC connectivity, implemented using ContainerLab and FRRouting (FRR), and incorporates ECMP routing, BFD, and a queue-pair-aware traffic distribution mechanism. The framework emulates realistic WAN conditions to characterize communication and system behavior under AllReduce and Parameter Server workloads, providing insights into traffic distribution, resilience, and infrastructure choices while maintaining compatibility with commodity hardware.

Significance. If the emulation accurately captures real WAN conditions and the reported behaviors hold, the work could supply a reproducible, commodity-compatible platform for systematically evaluating infrastructure decisions in geo-distributed AI training, an area of growing importance due to model scale and data sovereignty constraints. The emphasis on EVPN-VXLAN, ECMP/BFD, and queue-pair mechanisms offers a practical bridge between networking research and synchronization-heavy AI workloads.

major comments (2)
  1. [Abstract] Abstract: The central claim that the framework yields actionable insights on traffic distribution and resilience under AllReduce and Parameter Server patterns rests on 'realistic WAN emulation,' yet the abstract (and by extension the manuscript) supplies no quantitative metrics, error bars, baseline comparisons, or validation of emulated latency/jitter/loss distributions against production inter-DC traces or hardware testbeds.
  2. [Implementation/Evaluation (assumed §4–§5)] Implementation and evaluation sections: The assertion that ContainerLab plus FRRouting faithfully reproduces WAN latency, jitter, packet-loss, and correlation structure for scale AllReduce/PS workloads is load-bearing for all downstream claims about communication behavior and infrastructure conclusions, but no validation experiments (e.g., statistical comparison to real WAN traces) are described.
minor comments (2)
  1. [Abstract] The abstract is lengthy and could be condensed to foreground the specific quantitative findings once they are added.
  2. Notation for the queue-pair-aware distribution mechanism should be defined more explicitly if equations or pseudocode are introduced later in the manuscript.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on the validation of the WAN emulation in ScaleAcross. We address each major comment below and will revise the manuscript accordingly.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central claim that the framework yields actionable insights on traffic distribution and resilience under AllReduce and Parameter Server patterns rests on 'realistic WAN emulation,' yet the abstract (and by extension the manuscript) supplies no quantitative metrics, error bars, baseline comparisons, or validation of emulated latency/jitter/loss distributions against production inter-DC traces or hardware testbeds.

    Authors: We agree that the abstract does not provide quantitative validation metrics or comparisons for the emulated WAN conditions. The framework implements configurable emulation of latency, jitter, and loss via ContainerLab and FRR parameters drawn from typical inter-DC values in the literature, but no statistical validation against production traces is present. We will revise the abstract to describe the emulation more precisely as 'configurable WAN emulation' rather than 'realistic,' qualify the resulting insights accordingly, and add a dedicated subsection in the evaluation section that documents the chosen parameters, their grounding, and the absence of direct trace-based validation as a limitation. revision: yes

  2. Referee: [Implementation/Evaluation (assumed §4–§5)] Implementation and evaluation sections: The assertion that ContainerLab plus FRRouting faithfully reproduces WAN latency, jitter, packet-loss, and correlation structure for scale AllReduce/PS workloads is load-bearing for all downstream claims about communication behavior and infrastructure conclusions, but no validation experiments (e.g., statistical comparison to real WAN traces) are described.

    Authors: We acknowledge that the manuscript asserts the use of realistic WAN emulation without including validation experiments or statistical comparisons to real traces. This is a substantive gap for claims about communication behavior. In revision we will expand the implementation and evaluation sections to include (where feasible) additional analysis or experiments that characterize emulation fidelity, or else add an explicit limitations discussion on the emulation approach and its grounding in standard tooling and literature parameters. revision: yes

Circularity Check

0 steps flagged

No circularity: engineering framework description with no derivations or fitted predictions

full rationale

The paper describes an implemented emulation framework (VXLAN+EVPN via ContainerLab+FRR, with ECMP/BFD and queue-pair mechanisms) and reports observed behavior under AllReduce/Parameter Server workloads. No equations, parameters fitted to data subsets, predictions, or self-citation chains appear in the provided text. The central claims rest on emulation results rather than any reduction to inputs by construction. The noted limitation (unvalidated WAN fidelity) is an external-validity concern, not circularity. This matches the default case of a self-contained systems paper.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review supplies no equations, fitted constants, or explicit assumptions beyond standard networking primitives; therefore the ledger is empty.

pith-pipeline@v0.9.1-grok · 5748 in / 1104 out tokens · 19650 ms · 2026-06-27T05:39:55.830616+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

40 extracted references · 14 canonical work pages · 1 internal anchor

  1. [1]

    Hedera:dynamicflowschedulingfordatacenternetworks., in: Nsdi, San Jose, USA

    Al-Fares,M.,Radhakrishnan,S.,Raghavan,B.,Huang,N.,Vahdat,A.,etal.,2010. Hedera:dynamicflowschedulingfordatacenternetworks., in: Nsdi, San Jose, USA. pp. 89–92

    2010

  2. [2]

    Conga: Distributed congestion-aware load balancing for datacenters, in: Proceedings of the 2014 ACM conference on SIGCOMM, pp

    Alizadeh, M., Edsall, T., Dharmapurikar, S., Vaidyanathan, R., Chu, K., Fingerhut, A., Lam, V.T., Matus, F., Pan, R., Yadav, N., et al., 2014. Conga: Distributed congestion-aware load balancing for datacenters, in: Proceedings of the 2014 ACM conference on SIGCOMM, pp. 503– 514

    2014

  3. [3]

    The evolution of the carrier cloud networking, in: 2013 IEEE Seventh International Symposium on Service- Oriented System Engineering, IEEE

    Cai, D., Natarajan, S., 2013. The evolution of the carrier cloud networking, in: 2013 IEEE Seventh International Symposium on Service- Oriented System Engineering, IEEE. pp. 286–291

    2013

  4. [4]

    Multiprotocol Extensions for BGP-4

    Chandra, R., Bates, T.J., Rekhter, Y., Katz, D., 2007. Multiprotocol Extensions for BGP-4. RFC 4760. URL:https://www.rfc-editor. org/info/rfc4760, doi:10.17487/RFC4760. accessed: 2025-10-21

    work page doi:10.17487/rfc4760 2007

  5. [5]

    containerlab: Container-based networking lab framework.https://github.com/srl-labs/containerlab

    Dodin, R., et al., 2020. containerlab: Container-based networking lab framework.https://github.com/srl-labs/containerlab. Accessed: 2025-06-10

    2020

  6. [6]

    Beyondasingleaicluster:Asurveyofdecentralizedllmtraining

    Dong,H.,Jiang,J.,Lu,R.,Luo,J.,Song,J.,Li,B.,Shen,Y.,Wang,Z.,2025. Beyondasingleaicluster:Asurveyofdecentralizedllmtraining. arXiv preprint arXiv:2503.11023

    work page arXiv 2025

  7. [7]

    T., Zaman, A., Sajid, M

    Elmadani, M., Sati, S.O., 2023. Data center lab using vxlan data plane and bgp-evpn control plane, in: 2023 4th International Conference on Data Analytics for Business and Industry (ICDABI), pp. 354–358. doi:10.1109/ICDABI60145.2023.10629438

    work page doi:10.1109/icdabi60145.2023.10629438 2023

  8. [8]

    Rdmaoverethernet for distributed training at meta scale, in: Proceedings of the ACM SIGCOMM 2024 Conference, pp

    Gangidi,A.,Miao,R.,Zheng,S.,Bondu,S.J.,Goes,G.,Morsy,H.,Puri,R.,Riftadi,M.,Shetty,A.J.,Yang,J.,etal.,2024. Rdmaoverethernet for distributed training at meta scale, in: Proceedings of the ACM SIGCOMM 2024 Conference, pp. 57–70

    2024

  9. [9]

    Scc: Synchronization congestion control for multi-tenant learning over geo-distributed clouds

    Gao, C., Li, F., Ye, K., Wang, Y., Wang, P., Wang, X., Xu, C., 2025. Scc: Synchronization congestion control for multi-tenant learning over geo-distributed clouds. IEEE Transactions on Computers

    2025

  10. [10]

    A survey of virtual private lan services (vpls): Past, present and future

    Gaur, K., Kalla, A., Grover, J., Borhani, M., Gurtov, A., Liyanage, M., 2021. A survey of virtual private lan services (vpls): Past, present and future. Computer Networks 196, 108245

    2021

  11. [11]

    Ge,H.,Feng,J.,Huang,Q.,Fu,F.,Nie,X.,Zuo,L.,Lin,H.,Cui,B.,Liu,X.,2025. Bytescale:Communication-efficientscalingofllmtraining witha2048kcontextlengthon16384gpus,in:ProceedingsoftheACMSIGCOMM2025Conference,AssociationforComputingMachinery, New York, NY, USA. p. 963–978. URL:https://doi.org/10.1145/3718958.3754352, doi:10.1145/3718958.3754352

    work page doi:10.1145/3718958.3754352 2025

  12. [12]

    Evolution of data center design to handle ai workloads, in: 2024 34th International Telecommunication Networks and Applications Conference (ITNAC), IEEE

    Gupta, T., 2024. Evolution of data center design to handle ai workloads, in: 2024 34th International Telecommunication Networks and Applications Conference (ITNAC), IEEE. pp. 1–8

    2024

  13. [13]

    Reproduciblenetworkexperimentsusingcontainer-basedemulation, in: Proceedings of the 8th international conference on Emerging networking experiments and technologies, pp

    Handigol,N.,Heller,B.,Jeyakumar,V.,Lantz,B.,McKeown,N.,2012. Reproduciblenetworkexperimentsusingcontainer-basedemulation, in: Proceedings of the 8th international conference on Emerging networking experiments and technologies, pp. 253–264

    2012

  14. [14]

    Hoefler,T.,Roweth,D.,Underwood,K.,Alverson,R.,Griswold,M.,Tabatabaee,V.,Kalkunte,M.,Anubolu,S.,Shen,S.,McLaren,M.,etal.,

  15. [15]

    Computer 56, 67–77

    Data center ethernet and remote direct memory access: Issues at hyperscale. Computer 56, 67–77

  16. [16]

    Hoefler, T., Schramm, K., Spada, E., Underwood, K., Alexander, C., Alverson, B., Bottorff, P., Caulfield, A., Handley, M., Huang, C., et al.,

  17. [17]

    Ultra ethernet’s design principles and architectural innovations,

    Ultra ethernet’s design principles and architectural innovations. arXiv preprint arXiv:2508.08906

    work page arXiv

  18. [18]

    Analysis of an equal-cost multi-path algorithm

    Hopps, C., 2000. Analysis of an equal-cost multi-path algorithm. Technical Report

    2000

  19. [19]

    L3dml: Facilitating geo-distributed machine learning in network layer

    Hou, X., Gao, S., Liu, N., Yao, F., Lei, B., Zhang, H., Das, S., 2024. L3dml: Facilitating geo-distributed machine learning in network layer. IEEE Transactions on Network and Service Management PP, 1–1. doi:10.1109/TNSM.2024.3509031

    work page doi:10.1109/tnsm.2024.3509031 2024

  20. [20]

    Gaia: Geo-Distributed machine learning approachingLANspeeds,in:14thUSENIXSymposiumonNetworkedSystemsDesignandImplementation(NSDI17),USENIXAssociation, Boston, MA

    Hsieh, K., Harlap, A., Vijaykumar, N., Konomis, D., Ganger, G.R., Gibbons, P.B., Mutlu, O., 2017. Gaia: Geo-Distributed machine learning approachingLANspeeds,in:14thUSENIXSymposiumonNetworkedSystemsDesignandImplementation(NSDI17),USENIXAssociation, Boston, MA. pp. 629–647. URL:https://www.usenix.org/conference/nsdi17/technical-sessions/presentation/hsieh....

    2017

  21. [21]

    Demystifyingnccl:Anin-depth analysis of gpu communication protocols and algorithms, in: 2025 IEEE Symposium on High-Performance Interconnects (HOTI), IEEE

    Hu,Z.,Shen,S.,Bonato,T.,Jeaugey,S.,Alexander,C.,Spada,E.,Dinan,J.,Hammond,J.,Hoefler,T.,2025. Demystifyingnccl:Anin-depth analysis of gpu communication protocols and algorithms, in: 2025 IEEE Symposium on High-Performance Interconnects (HOTI), IEEE. pp. 48–59

    2025

  22. [22]

    Jiang, Z., Lin, H., Zhong, Y., Huang, Q., Chen, Y., Zhang, Z., Peng, Y., Li, X., Xie, C., Nong, S., et al., 2024.{MegaScale}: Scaling large language model training to more than 10,000{GPUs}, in: 21st USENIX Symposium on Networked Systems Design and Implementation (NSDI 24), pp. 745–760

    2024

  23. [23]

    Bidirectional Forwarding Detection (BFD)

    Katz, D., Ward, D., 2010. Bidirectional Forwarding Detection (BFD). RFC 5880. URL:https://www.rfc-editor.org/rfc/rfc5880. html, doi:10.17487/RFC5880. accessed: 2025-10-21

    work page doi:10.17487/rfc5880 2010

  24. [24]

    Parallelgradientcomputationandsynchronization:Enhancingtheefficiencyofdistributed training for llms

    Li,H.,Jiang,H.,Wu,J.,Yang,G.,Zhang,J.,2025. Parallelgradientcomputationandsynchronization:Enhancingtheefficiencyofdistributed training for llms. IEEE Transactions on Network Science and Engineering

    2025

  25. [25]

    Communicationefficientdistributedmachinelearningwiththeparameterserver

    Li,M.,Andersen,D.G.,Smola,A.,Yu,K.,2014. Communicationefficientdistributedmachinelearningwiththeparameterserver. Advances in neural information processing systems 27

    2014

  26. [26]

    arXiv preprint arXiv:2404.06114

    Liang,F.,Zhang,Z.,Lu,H.,Leung,V.,Guo,Y.,Hu,X.,2024.Communication-efficientlarge-scaledistributeddeeplearning:Acomprehensive survey. arXiv preprint arXiv:2404.06114

    work page arXiv 2024

  27. [27]

    Virtual eXtensible Local Area Network (VXLAN): A Framework for Overlaying Virtualized Layer 2 Networks over Layer 3 Networks

    Mahalingam, M., Dutt, D., Duda, K., Agarwal, P., Kreeger, L., Sridhar, T., Bursell, M., Wright, C., 2014. Virtual eXtensible Local Area Network (VXLAN): A Framework for Overlaying Virtualized Layer 2 Networks over Layer 3 Networks. RFC 7348. URL:https: //www.rfc-editor.org/info/rfc7348, doi:10.17487/RFC7348. accessed: 2025-10-21

    work page doi:10.17487/rfc7348 2014

  28. [28]

    Federated learning of deep networks using model averaging,

    McMahan, H.B., Moore, E., Ramage, D., Hampson, S., y Arcas, B.A., 2023. Communication-efficient learning of deep networks from decentralized data. URL:https://arxiv.org/abs/1602.05629,arXiv:1602.05629

    work page arXiv 2023

  29. [29]

    Collaborative deep learning across multiple data centers

    Mi, H., Xu, K., Feng, D., Wang, H., Zhang, Y., Zheng, Z., Chen, C., Lan, X., 2020. Collaborative deep learning across multiple data centers. Science China Information Sciences 63, 182102

    2020

  30. [30]

    Miao, X., Nie, X., Shao, Y., Yang, Z., Jiang, J., Ma, L., Cui, B., 2021. Heterogeneity-aware distributed machine learning training via partial reduce,in:Proceedingsofthe2021InternationalConferenceonManagementofData,AssociationforComputingMachinery,NewYork,NY, USA. p. 2262–2270. URL:https://doi.org/10.1145/3448016.3452773, doi:10.1145/3448016.3452773

    work page doi:10.1145/3448016.3452773 2021

  31. [31]

    Evpn/sdn assisted live vm migration between geo-distributed data centers, in: 2018 4th IEEE Conference on Network Softwarization and Workshops (NetSoft), pp

    Noghani, K.A., Kassler, A., Gopannan, P.S., 2018. Evpn/sdn assisted live vm migration between geo-distributed data centers, in: 2018 4th IEEE Conference on Network Softwarization and Workshops (NetSoft), pp. 105–113. doi:10.1109/NETSOFT.2018.8459946

    work page doi:10.1109/netsoft.2018.8459946 2018

  32. [32]

    accessed: 2025-06-08

    NVIDIADeveloperBlog,2024.Turbochargellmtrainingacrosslong-hauldatacenternetworkswithnvidianemoframework.URL:https:// developer.nvidia.com/blog/turbocharge-llm-training-across-long-haul-data-center-networks-with-nvidia-nemo-framework/. accessed: 2025-06-08

    2024

  33. [33]

    Datacenter optimization methods for softwarized network services

    Pannocchi, L., Lahiri, S., Fichera, S., Artale, A., Cucinotta, T., 2024. Datacenter optimization methods for softwarized network services. Journal of Systems Architecture 156, 103270

    2024

  34. [34]

    The ns-3 network simulator, in: Modeling and tools for network simulation

    Riley, G.F., Henderson, T.R., 2010. The ns-3 network simulator, in: Modeling and tools for network simulation. Springer, pp. 15–34

    2010

  35. [35]

    Acceleratingcollectivecommunicationindataparalleltrainingacrossdeeplearningframeworks,in:19thUSENIXSymposiumonNetworked Systems Design and Implementation (NSDI 22), pp

    Romero, J., Yin, J., Laanait, N., Xie, B., Young, M.T., Treichler, S., Starchenko, V., Borisevich, A., Sergeev, A., Matheson, M., 2022. Acceleratingcollectivecommunicationindataparalleltrainingacrossdeeplearningframeworks,in:19thUSENIXSymposiumonNetworked Systems Design and Implementation (NSDI 22), pp. 1027–1040

    2022

  36. [36]

    A Network Virtualization Overlay Solution Using Ethernet VPN (EVPN)

    Sajassi, A., Drake, J., Bitar, N., Shekhar, R., Uttaro, J., Henderickx, W., 2018. A Network Virtualization Overlay Solution Using Ethernet VPN (EVPN). RFC 8365. URL:https://www.rfc-editor.org/info/rfc8365, doi:10.17487/RFC8365. accessed: 2025-10-21

    work page doi:10.17487/rfc8365 2018

  37. [37]

    librxe-dev: Software rdma over ethernet (soft-roce) implementation.https://github.com/SoftRoCE/ librxe-dev

    Soft-RoCE Development Team, . librxe-dev: Software rdma over ethernet (soft-roce) implementation.https://github.com/SoftRoCE/ librxe-dev. Accessed: 2026-05-20

    2026

  38. [38]

    Hcec:Anefficientgeo-distributeddeeplearningtrainingstrategybasedonwait-free back-propagation

    Song,Y.,Ai,Y.,Xiao,X.,Liu,Z.,Tang,Z.,Li,K.,2024. Hcec:Anefficientgeo-distributeddeeplearningtrainingstrategybasedonwait-free back-propagation. Journal of Systems Architecture 148, 103070

    2024

  39. [39]

    Won,W.,Heo,T.,Rashidi,S.,Sridharan,S.,Srinivasan,S.,Krishna,T.,2023.Astra-sim2.0:Modelinghierarchicalnetworksanddisaggregated systemsforlarge-modeltrainingatscale,in:2023IEEEInternationalSymposiumonPerformanceAnalysisofSystemsandSoftware(ISPASS), IEEE. pp. 283–294

    2023

  40. [40]

    Collaborative Deep Learning Across Multiple Data Centers

    Xu, K., Mi, H., Feng, D., Wang, H., Chen, C., Zheng, Z., Lan, X., 2018. Collaborative deep learning across multiple data centers. URL: https://arxiv.org/abs/1810.06877,arXiv:1810.06877. :Preprint submitted to Elsevier Page 19 of 19

    work page internal anchor Pith review Pith/arXiv arXiv 2018