Pith. sign in

REVIEW 3 major objections 4 minor 2 cited by

OpenOptics: An Open Research Framework for Optical Data Center Networks

T0 review · 3 major / 4 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read A time-flow table abstracts all optical data center routing, cutting circuit duration to 2 µs.

desk verdict A real systems contribution with a genuinely new abstraction and credible testbed validation, but the headline 2 µs record leans on an unpublished sync protocol that needs to be checked. read the letter →

arxiv 2411.18319 v2 pith:XQRSM54H submitted 2024-11-27 cs.NI

Yiming Lei , Federico De Marchi , Jialong Li , Raj Joshi , Balakrishnan Chandrasekaran , Yiting Xia This is my paper · ORCID
classification cs.NI
keywords opticaldatacenternetworkstime-flowtableprogrammableswitchestraffic-awareroutingtraffic-obliviouscircuitswitchingqueuemanagementopenresearchframework
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

OpenOptics claims that optical data center networks do not need bespoke hardware-software pairings. The paper introduces a time-flow table, a routing abstraction in which a packet's arrival time slice is a match field and its departure time slice is an action field, and argues that this single abstraction can express traffic-aware, traffic-oblivious, and traditional static topologies. On top of it, a queue-management backend reassigns switch egress queues to future time slices and pauses and resumes them in synchrony, enabling microsecond-scale circuit use on commodity programmable switches. The paper validates the framework by implementing six published optical architectures and seven routing schemes, and it reports a minimum optical circuit duration of 2 µs from a guardband analysis. A sympathetic reader would care because this is a path from closed, hard-to-reproduce optical data center proposals to a common research platform where architectures can be compared and extended.

What carries the argument

The load-bearing objects are the time-flow table and the calendar-queue rotation mechanism. The time-flow table uses arrival time slice as a match field and departure time slice as an action field, so a packet can be buffered and sent in a later slice or forwarded immediately when the slices coincide, and with wildcards it degenerates to a legacy flow table. The backend assigns one calendar queue per time slice on each egress port, pauses every queue except the active one, and rotates the active queue at each slice boundary using a packet generator synchronized to the optical controller; this turns a switch's limited buffer into time-aware storage. The same machinery supports congestion detection through estimated queue occupancy, traffic push-back, flow pausing, and buffer offloading.

What would settle it

Measure the synchronization error between two top-of-rack switches and the optical controller while running the 2 µs time slice under production traces; if the observed error exceeds the 56 ns bidirectional guardband or the queue rotation variance exceeds 34 ns, packet loss should appear during circuit transitions.

Watch

Extended reading notes

Core claim

The central claim is that the time-flow table is a sufficient narrow waist for optical data center networks: every routing primitive used by existing optical architectures, including direct-circuit, multi-hop, source, and multi-path routing, can be represented as entries that match on arrival time slice and act on departure time slice and egress port, and the same table reduces to an ordinary flow table when both time fields are wildcarded. The paper further claims that commodity programmable switches can execute this abstraction at 2 µs time slices by rotating calendar queues: each egress port maintains queues assigned to future time slices, an on-chip packet generator triggers queue rotation at each slice boundary, and nanosecond-scale synchronization aligns switches, NICs, and the optical controller. Evidence includes implementation of six architectures and seven routing schemes, a 108-ToR benchmark with hardware resource usage under 13.8%, and a 200 ns guardband decomposed into synchronization error, queue rotation variance, and occupancy estimation error.

Load-bearing premise

The framework relies on a nanosecond-precision synchronization protocol, described in an unpublished companion paper, to align switches, NICs, and optical controller; if that protocol is unavailable or its 28 ns error is optimistic, the 2 µs minimum slice and the queue-rotation design do not hold.

Editorial extensions

If this is right

  • If the claims hold, researchers can implement a new optical data center architecture by writing a short Python program that calls topology and routing APIs, instead of building a custom controller and flow tables.
  • Traffic-oblivious designs such as RotorNet and Opera and traffic-aware designs such as Jupiter, Mordia, and c-Through can be compared on the same testbed, which the paper demonstrates with flow completion time distributions.
  • The 2 µs minimum slice implies that microsecond-scale optical proposals, previously only simulated, are realizable on commodity equipment with the stated guardband.
  • Hybrid traffic-aware plus traffic-oblivious designs become directly programmable, including semi-oblivious schedules and hierarchical combinations of scale-up and scale-out networks.
  • Resource usage under 13.8% on commodity switch hardware suggests headroom for larger networks and more complex routing, though the paper does not demonstrate a full 108-ToR deployment.

Reading between the lines

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

  • The guardband decomposition suggests a scaling law: if synchronization error or queue rotation variance were halved, the minimum slice could approach 1 µs, because the 148 ns guardband is dominated by measured system components rather than by OCS hardware limits.
  • The time-flow table could be ported to host-centric NICs and other programmable dataplanes, but the paper reports that a host version is under development, so generality beyond switch-centric designs is not yet demonstrated.
  • A testable extension is to run the same six-architecture comparison with a real OCS reconfigured at sub-millisecond rates rather than an emulated fabric, to separate emulation fidelity from the claimed 2 µs limit.
  • Another inference is that the queue-occupancy estimation error of 725 bytes bounds how small a time slice can become before congestion detection becomes unreliable, so improving estimation accuracy would likely shift the achievable minimum slice lower.
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 / 4 minor

Summary. OpenOptics is a framework for optical DCN research built on Intel Tofino2 programmable switches and Mellanox NICs. It proposes a time-flow table abstraction that adds arrival and departure time-slice match/action fields to traditional flow tables, a Python API and compiler workflow that unify traffic-aware and traffic-oblivious architecture implementations, and a switch/NIC backend with calendar-queue-based time scheduling, estimated queue occupancy, congestion detection, push-back, flow pausing, and buffer offloading. The system is evaluated on a testbed with a Polatis MEMS OCS, four Tofino2 switches, and four servers; six architectures and seven routing schemes are implemented, and a 108-ToR-scale benchmark is run with one physical ToR and an emulated fabric. The paper claims a record minimum optical circuit duration of 2 us and validates behavior by reproducing qualitative FCT/RTT trends from prior work.

Significance. If the claims hold, OpenOptics is a valuable community resource: it is the first system, to my knowledge, to implement a general time-flow table abstraction on commodity programmable switches, to unify TA and TO workflows in one control plane, and to demonstrate microsecond-scale time slices on commodity hardware. The authors' decision to run on a real testbed with a physical MEMS OCS, and to include an emulated-fabric and Mininet path, strengthens the paper's reproducibility ambitions. The main significance caveat is that the 2 us record and the queue-rotation mechanism rely on a nanosecond synchronization protocol that is only cited as an anonymous companion paper, so the central quantitative claim is not yet self-contained. The qualitative validation against RotorNet and the detailed resource-usage measurements are useful, but they do not fully substitute for a quantitative reproduction or an available artifact.

major comments (3)
  1. [§7 (Minimum time slice duration), §4.1, §5.1] The 2 us figure is computed as 10× a 200 ns guardband whose largest identified component (56 ns of the 148 ns total) is the 28 ns synchronization error cited to ref [1], which the manuscript itself describes as 'a separate paper' (§4.1 and §5.1) and which is listed as 'Anonymous time synchronization paper' in the references. Because the queue-rotation mechanism in §5.1 assumes all switches, NICs, and the optical controller share slice boundaries at nanosecond precision, this term is not an incidental detail: both the headline record and the mechanism's correctness depend on an artifact the reader cannot inspect. The paper should either include a self-contained description and measurement of the synchronization protocol, or explicitly qualify the 2 us claim as conditional on [1] becoming available.
  2. [§7 (Experimental setup)] The 108-ToR benchmark is executed with one Tofino2 as the observed ToR, a second Tofino2 as the emulated optical fabric, and six hosts; the full 108-ToR time-flow table is populated only on that single ToR. This measures per-ToR resource usage and emulated-fabric behavior, but it does not exercise multi-ToR synchronization, controller coordination across many ToRs, or the interaction of calendar-queue rotation across multiple switches. Consequently, statements in the abstract and conclusion about benchmarking 'a 108-ToR setup' and drawing scalability conclusions from Table 2 overstate what the experiment can show. Please rephrase to 'single-ToR benchmark with a 108-ToR-scale table and emulated fabric' and temper the scaling claims.
  3. [§6 (Testbed) and §7 (Emulation accuracy)] Correctness validation relies on qualitative shape matches: FCT trends in Fig. 8 are said to 'match the FCT trends reported in the concerned prior work,' and Fig. 13 shows 'similar curve shapes' with lower absolute RTTs and no tail compared with RotorNet. The paper acknowledges it is 'inferring performance trends' rather than reproducing published numbers. For the claim that OpenOptics correctly implements six architectures, this is a weak form of evidence; a quantitative comparison (e.g., matching published FCT/RTT distributions under identical parameters, or providing a reproducibility artifact) would make the central correctness claim more load-bearing. This is not an internal inconsistency, but it should be addressed before the correctness claims are accepted as stated.
minor comments (4)
  1. [§7, Fig. 12] The sentence 'The estimation accuracy increases with the update interval' appears inverted: Fig. 12 shows larger estimation errors for longer update intervals (100 ns, 150 ns) than for shorter ones, so the accuracy decreases as the update interval increases. Please correct the wording.
  2. [§7 (Experimental setup), references] Reference [36] is cited for the RPC trace in the experimental setup, but [36] is the Homa paper; the RPC trace appears to be the Facebook/RPC trace from [40]/[41]. In addition, references [40] and [41] are duplicate entries with identical bibliographic data. Please fix the citations.
  3. [Abstract and §7] The term 'record-breaking' is not placed in context: the comparison set for 'the lowest ever achieved with commodity network devices' is not defined, and no prior minimum time-slice durations are cited. Please state the prior art against which 2 us is a record.
  4. [§5.2 (Congestion detection)] The text says the 50 ns update interval results in 'less than one packet estimation error' and cites 725 B from Fig. 12; this is less than one 1500 B MTU packet but more than one 60 B packet. Please specify the comparison packet size.

Circularity Check

1 steps flagged · score 4.0 of 10

Headline 2 us record rests on an unverified self-cited synchronization-error figure; the core framework is otherwise independently benchmarked.

  1. self citation load bearing [Sec. 7, 'Minimum time slice duration'; also Secs. 4.1 and 5.1]
    "Besides, our synchronization work shows up to 28 ns sync errors in a 192-ToR optical DCN [1], which requires a guardband of 28×2 = 56 ns for clock discrepancies above and below the actual clock. Therefore, the total guardband is 34+58+56 = 148 ns, and with added headroom for runtime variations, we set the guardband to 200 ns. OpenOptics thus supports a minimum time slice duration of 200 ns×10 = 2 µs"

    The record-breaking 2 us minimum slice is not measured end-to-end; it is computed as 10x the 200 ns guardband. 56 of the 148 ns guardband constituents (28 ns x 2) come from ref. [1], which the paper describes as 'our synchronization work.' This same nanosecond-precision sync is also the correctness precondition for the queue-rotation design in Sec. 5.1. Ref. [1] is an anonymous, unpublished companion paper; no protocol, measurement, or reproducibility evidence is included here, so the headline number's quantitative core reduces to an unverifiable self-citation. It is load-bearing rather than a minor citation, though the 34 ns and 58 ns terms are independently measured in Figs. 11 and 12.

full rationale

OpenOptics' main contributions--the time-flow table abstraction, the unified API, the queue-management rearchitecture, and the emulated-fabric design--are derived from the system's own design choices and are validated against independent external references: FCT trends from prior implemented architectures (RotorNet, Opera, c-Through, Jupiter, Mordia) and a reproduced RotorNet UDP-latency experiment. The 34 ns queue-rotation variance and 58 ns occupancy-estimation contribution to the guardband are measured in this paper (Figs. 11 and 12). The only load-bearing step that routes through the authors' own unpublished work is the 28 ns synchronization error from ref. [1], which contributes 56 ns of the 148 ns guardband and hence scales linearly into the claimed 2 us minimum slice duration. That is a dependency and a missing-support concern, not a definitional tautology: the 2 us figure would change if the actual sync error were different, and it is not statistically forced by any parameter fitted in this paper. However, because the central quantitative claim depends materially on a self-citation that is not independently checkable from the preprint, a moderate score of 4 is appropriate rather than 0. The framework's independent content keeps this from being a fully circular derivation.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claims rest on five stated premises (sync accuracy, line-rate dequeuing, emulation fidelity, the 10x guardband rule, and single-ToR representativeness) plus four hand-chosen quantities (guardband headroom, duty-cycle factor, update interval, and benchmark slice). No new physical entities are postulated; the time-flow table is a software abstraction implemented in the system.

free parameters (4)
  • Guardband headroom = 200 ns total guardband
    Section 7: measured components sum to 148 ns, then 'with added headroom for runtime variations, we set the guardband to 200 ns'. Hand-chosen margin that directly sets the 2 µs minimum slice.
  • Duty-cycle factor = 10
    Section 7: 'to ensure a duty cycle above 90%, the slice duration is typically set to at least 10x the guardband.' Design rule, not derived; 2 µs = 10 x 200 ns.
  • Queue occupancy update interval = 50 ns
    Appendix A: 'We find 50 ns draws a good balance between estimation accuracy and pipeline processing.' Tuned on the testbed; estimation error stays below one packet.
  • Benchmark time slice duration = 300 µs
    Appendix A and B: chosen as 'considered long' for TO architectures for buffer usage and congestion tests; not a fixed system parameter.
assumptions (5)
  • ad hoc to paper Nanosecond-precision time synchronization among switches, NICs, and the optical controller is achievable and available (cited to unpublished ref [1]).
    Sections 4.1 and 5.1 rely on this for queue rotation and time-flow table scheduling; the paper defers the protocol to an anonymous companion paper, so this premise is not self-contained.
  • domain assumption Egress queues dequeue at line rate so occupancy registers can be decremented by bandwidth times the update interval.
    Appendix A assumes line-rate dequeuing to estimate queue occupancy; under congestion or bursty traffic this may over- or under-estimate occupancy.
  • domain assumption The P4-based emulated OCS with cut-through mode closely approximates physical OCS behavior.
    Section 5.3 and the 2 µs minimum slice analysis rely on the emulated fabric; only one comparison to a real OCS (RotorNet, Fig. 13) is provided.
  • ad hoc to paper A 10x guardband-to-slice ratio (duty cycle above 90%) is a valid design rule for optical DCN operation.
    Section 7 'Minimum time slice duration' asserts this as typical practice rather than deriving it; the 2 µs result scales directly from this choice.
  • domain assumption A single physical Tofino2 ToR with a second Tofino2 emulating the fabric is representative of the full 108-ToR network.
    Section 7 populates full time-flow table entries and replays traces on one observed ToR; inter-ToR interactions and end-to-end congestion at scale are not directly measured.

how reviews work

0 comments
Cite this review

Pith. "Pith review of OpenOptics: An Open Research Framework for Optical Data Center Networks." pith.science (2026). https://pith.science/paper/XQRSM54H

@misc{pith2026241118319,
  author       = {Pith},
  title        = {Pith review of: OpenOptics: An Open Research Framework for Optical Data Center Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XQRSM54H}},
  note         = {Machine review of arXiv:2411.18319}
}
abstract

Optical data center networks (DCNs) are emerging as a promising design for cloud infrastructure. However, existing optical DCN architectures operate as closed ecosystems, tying software solutions to specific optical hardware. We introduce OpenOptics, an open research framework that decouples software from hardware, allowing them to evolve independently. OpenOptics features: (1) a time-flow table abstraction as a common interface between optical hardware and software, (2) a unified workflow and user-friendly API for implementing various optical DCNs with simple Python scripts, and (3) a backend system that re-architects queue management to support the time-flow tables and provides rich infrastructure services for diverse applications. Built on programmable switches, OpenOptics achieves a record-breaking minimum optical circuit duration of 2 $\mu$s using commodity devices. We validate OpenOptics' generality by implementing six optical architectures and seven routing schemes on an optical testbed and conducting benchmarks on a 108-ToR setup, showcasing its efficiency. Additionally, case studies highlight novel research opportunities enabled by OpenOptics.

Figures

Figures reproduced from arXiv: 2411.18319 by the authors.

Figure 1
Figure 1. An example optical DCN and its three topologies connected by different optical circuits over the OCSes. The tight coupling of optical hardware and software systems makes research and innovation for optical DCNs siloed. Each optical DCN architecture is a closed ecosystem comprising specialized optical hardware and customized networked sys￾tems to support that hardware. System solutions are tied to the underlying hard… view at source ↗
Figure 2
Figure 2. A routing example in a TO optical DCN. The arrows represent the paths for a packet from endpoint node 𝑁1 to 𝑁3 arriving at time slice 𝑡𝑠=0 under different routing schemes. the optical DCN as a static network with occasional topology updates adapting to long-term traffic patterns [16, 17, 48]. Google’s Jupiter and Lightwave fabrics follow this approach, with reconfiguration intervals of minutes to hours [31, 39]. TO … view at source ↗
Figure 3
Figure 3. Time-flow table examples for: (a) direct path ○1 and (b) multi-hop path ○2 in [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: OpenOptics system workflow. are the emulated OCSes on a programmable switch, which offer users access without the need for real OCSes. We also implement the entire workflow and system in Mininet (omit￾ted from the figure) to make OpenOptics accessible to a broader rang…
Figure 5
Figure 5. Figure 5: Code snippets of implementing various optical architectures using the OpenOptics APIs (§4.2). 4.2 User API A user first creates an OpenOptics network object and then calls the topology, routing, and monitoring APIs listed in Tab. 1. The API calls accept optional argume…
Figure 6
Figure 6. Figure 6: Example of packet processing on a switch, assuming the active calendar queue is q=0 for time slice ts=0. 5.1 Time-Based Queue Management The time-flow table requires time-based packet scheduling to dispatch packets at the planned departure time slice, and buffering is …
Figure 7
Figure 7. Figure 7: Testbed setup (servers in 7a are omitted from 7b). the total traffic sent to each destination switch and report the statistics to the optical controller. Buffer offloading. Some multi-hop routing schemes con￾sume excessive buffer space at intermediate switches, such as…
Figure 8
Figure 8. Figure 8: FCTs of (a) Memcached and (b) Gloo allreduce. Clos (3/5 Dupacks) Direct-circuit (3/5 Dupacks) Hybrid (3/5 Dupacks) VLB (3/5 Dupacks) VLB (3/5 Dupacks) Hybrid (3/5 Dupacks) [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 9
Figure 9. Figure 9: (a) TCP throughput and (b) number of packet re￾ordering events with iperf traffic. We measure the throughput of long-lasting iperf3 flows between hosts on Clos, RotorNet with VLB and direct-circuit routing, and finally a hybrid version of RotorNet with 100Gbps bandwidt…
Figure 10
Figure 10. Figure 10: Mice flow FCTs on RotorNet with OCSes of different time slice durations, under (a)VLB and (b)UCMP routing. OpenOptics facilitates more informed decisions on optical hardware selection through its emulation capabilities. We sample four recently proposed OCS technologie…
Figure 13
Figure 13. Figure 13: UDP latency in OpenOptics vs. in RotorNet (Fig.15 blue curve in [34]) [PITH_FULL_IMAGE:figures/full_fig_p012_13.png]
Figure 14
Figure 14. Figure 14: RTT delay (blue) and RTT distance to the 100 µs interval (red). eventually fill up, causing losses. When both mechanisms are combined (column 3), push-back engages once the primary queue is full, and slice-miss detection handles in-flight traffic before senders react.…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Experimental Evaluation of an SDN Controller for Open Optical-circuit-switched Networks

    cs.NI 2025-01 conditional novelty 6.0 of 10

    A multi-vendor SDN controller for optical circuit switches was implemented, and it configured fiber paths within 1.0 second on real and emulated networks.

  2. Unlocking Diversity of Fast-Switched Optical Data Center Networks with Unified Routing

    cs.NI 2024-11 conditional novelty 6.0 of 10

    A schedule-aware routing algorithm with in-switch buffering achieves loss-free operation in fast-switched optical data center networks at microsecond-scale time slices, shown on a Tofino2 prototype and in large simulations.

Reference graph

Works this paper leans on

48 extracted references · 45 canonical work pages · cited by 2 Pith papers

  1. [1]

    Anonymous time synchronization paper

    work page

  2. [2]

    https://github.com/p4lang/behavioral-model

    behavioral-model. https://github.com/p4lang/behavioral-model. Ac- cessed: 2025-1-21

    work page 2025

  3. [3]

    https://github.com/facebookincubator/gloo

    Gloo. https://github.com/facebookincubator/gloo

    work page

  4. [4]

    https://github.com/Mellanox/libvma/ blob/master/README

    Mellanox Messaging Accelerator. https://github.com/Mellanox/libvma/ blob/master/README

    work page

  5. [5]

    https://memcached.org/

    Memchached. https://memcached.org/

    work page

  6. [6]

    http://docs.libmemcached.org/bin/memslap.html

    Memslap. http://docs.libmemcached.org/bin/memslap.html

    work page

  7. [7]

    https://www.nsnam.org/

    NS3. https://www.nsnam.org/

    work page

  8. [8]

    https://omnetpp.org/

    OMNeT++. https://omnetpp.org/

    work page

Show all 48 references
  1. [9]

    https://opnetprojects.com/

    OPNET. https://opnetprojects.com/

  2. [10]

    A scalable, commodity data center network architecture.ACM SIGCOMM computer communication review, 38(4):63–74, 2008

    Mohammad Al-Fares, Alexander Loukissas, and Amin Vahdat. A scalable, commodity data center network architecture.ACM SIGCOMM computer communication review, 38(4):63–74, 2008

    2008

  3. [11]

    Shale: A practical, scalable oblivious reconfigurable network

    Daniel Amir, Nitika Saran, Tegan Wilson, Robert Kleinberg, Vishal Shrivastav, and Hakim Weatherspoon. Shale: A practical, scalable oblivious reconfigurable network. InProceedings of the ACM SIGCOMM 2024 Conference, pages 449–464, 2024

    2024

  4. [12]

    Workload analysis of a large-scale key-value store

    Berk Atikoglu, Yuehai Xu, Eitan Frachtenberg, Song Jiang, and Mike Paleczny. Workload analysis of a large-scale key-value store. In Proceedings of the 12th ACM SIGMETRICS/PERFORMANCE joint inter- national conference on Measurement and Modeling of Computer Systems, pages 53–64, 2012

    2012

  5. [13]

    Pias: Practical information-agnostic flow scheduling for commodity data centers

    Wei Bai, Li Chen, Kai Chen, Dongsu Han, Chen Tian, and Hao Wang. Pias: Practical information-agnostic flow scheduling for commodity data centers. IEEE/ACM Transactions on Networking, 25(4):1954–1967, 2017

    1954

  6. [14]

    Sirius: A flat datacenter network with nanosecond optical switching

    Hitesh Ballani, Paolo Costa, Raphael Behrendt, Daniel Cletheroe, Istvan Haller, Krzysztof Jozwik, Fotini Karinou, Sophie Lange, Kai Shi, Benn Thomsen, et al. Sirius: A flat datacenter network with nanosecond optical switching. In Proceedings of the Annual conference of the ACM...

    2020

  7. [15]

    Pulse: optical circuit switched data center ar- chitecture operating at nanosecond timescales

    Joshua L Benjamin, Thomas Gerard, Domaniç Lavery, Polina Bayvel, and Georgios Zervas. Pulse: optical circuit switched data center ar- chitecture operating at nanosecond timescales. Journal of Lightwave Technology, 38(18):4906–4921, 2020

    2020

  8. [16]

    Osa: An optical switching architecture for data center networks with unprecedented flexibility

    Kai Chen, Ankit Singla, Atul Singh, Kishore Ramachandran, Lei Xu, Yueping Zhang, Xitao Wen, and Yan Chen. Osa: An optical switching architecture for data center networks with unprecedented flexibility. IEEE/ACM Transactions on Networking, 22(2):498–511, 2013

    2013

  9. [17]

    Wavecube: A scalable, fault-tolerant, high- performance optical data center architecture

    Kai Chen, Xitao Wen, Xingyu Ma, Yan Chen, Yong Xia, Chengchen Hu, and Qunfeng Dong. Wavecube: A scalable, fault-tolerant, high- performance optical data center architecture. In 2015 IEEE Conference on Computer Communications (INFOCOM) , pages 1903–1911. IEEE, 2015

    2015

  10. [18]

    Enabling{Wide-Spread} communica- tions on optical fabric with{MegaSwitch}

    Li Chen, Kai Chen, Zhonghua Zhu, Minlan Yu, George Porter, Chun- ming Qiao, and Shan Zhong. Enabling{Wide-Spread} communica- tions on optical fabric with{MegaSwitch}. In 14th USENIX Symposium on Networked Systems Design and Implementation (NSDI 17), pages 577– 593, 2017

    2017

  11. [19]

    Time-division tcp for reconfigurable data center networks

    Shawn Shuoshuo Chen, Weiyang Wang, Christopher Canel, Srinivasan Seshan, Alex C Snoeren, and Peter Steenkiste. Time-division tcp for reconfigurable data center networks. In Proceedings of the ACM SIGCOMM 2022 Conference, pages 19–35, 2022

    2022

  12. [20]

    Helios: a hybrid electrical/optical switch architecture for modular data centers

    Nathan Farrington, George Porter, Sivasankar Radhakrishnan, Hamid Hajabdolali Bazzaz, Vikram Subramanya, Yeshaiahu Fainman, George Papen, and Amin Vahdat. Helios: a hybrid electrical/optical switch architecture for modular data centers. In Proceedings of the ACM SIGCOMM 2010 C...

    2010

  13. [21]

    Corundum: An open-source 100-gbps nic

    Alex Forencich, Alex C Snoeren, George Porter, and George Papen. Corundum: An open-source 100-gbps nic. In 2020 IEEE 28th Annual International Symposium on Field-Programmable Custom Computing Machines (FCCM), pages 38–46. IEEE, 2020

    2020

  14. [22]

    Dons: Fast and affordable discrete event network simulation with automatic parallelization

    Kaihui Gao, Li Chen, Dan Li, Vincent Liu, Xizheng Wang, Ran Zhang, and Lu Lu. Dons: Fast and affordable discrete event network simulation with automatic parallelization. In Proceedings of the ACM SIGCOMM 2023 Conference, pages 167–181, 2023

    2023

  15. [23]

    Projector: Agile re- configurable data center interconnect

    Monia Ghobadi, Ratul Mahajan, Amar Phanishayee, Nikhil Devanur, Ja- nardhan Kulkarni, Gireeja Ranade, Pierre-Alexandre Blanche, Houman Rastegarfar, Madeleine Glick, and Daniel Kilper. Projector: Agile re- configurable data center interconnect. In Proceedings of the 2016 ACM SI...

    2016

  16. [24]

    Timertasks: Towards time-driven execution in programmable dataplanes

    Raj Joshi, Ben Leong, and Mun Choon Chan. Timertasks: Towards time-driven execution in programmable dataplanes. In Proceedings of the ACM SIGCOMM 2019 Conference Posters and Demos , pages 69–71, 2019

    2019

  17. [25]

    Sip-ml: high-bandwidth optical network intercon- nects for machine learning training

    Mehrdad Khani, Manya Ghobadi, Mohammad Alizadeh, Ziyi Zhu, Madeleine Glick, Keren Bergman, Amin Vahdat, Benjamin Klenk, and Eiman Ebrahimi. Sip-ml: high-bandwidth optical network intercon- nects for machine learning training. In Proceedings of the 2021 ACM SIGCOMM 2021 Confere...

    2021

  18. [26]

    Advanced congestion & flow control with pro- grammable switches

    Jeongkeun Lee. Advanced congestion & flow control with pro- grammable switches. In P4 Expert Roundtable Series , 2020

    2020

  19. [27]

    Uniform-cost multi-path routing for recon- figurable data center networks

    Jialong Li, Haotian Gong, Federico De Marchi, Aoyu Gong, Yiming Lei, Wei Bai, and Yiting Xia. Uniform-cost multi-path routing for recon- figurable data center networks. In Proceedings of the ACM SIGCOMM 2024 Conference, pages 433–448, 2024

    2024

  20. [28]

    Hop-on hop-off routing: A fast tour across the optical data center network for latency-sensitive flows

    Jialong Li, Yiming Lei, Federico De Marchi, Raj Joshi, Balakrishnan Chandrasekaran, and Yiting Xia. Hop-on hop-off routing: A fast tour across the optical data center network for latency-sensitive flows. In Proceedings of the 6th Asia-Pacific Workshop on Networking , pages 63–69, 2022

    2022

  21. [29]

    Negotiator: Towards a simple yet effective on-demand reconfigurable datacenter network

    Cong Liang, Xiangli Song, Jing Cheng, Mowei Wang, Yashe Liu, Zhen- hua Liu, Shizhen Zhao, and Yong Cui. Negotiator: Towards a simple yet effective on-demand reconfigurable datacenter network. InProceedings of the ACM SIGCOMM 2024 Conference , pages 415–432, 2024

    2024

  22. [30]

    Circuit switching under the radar with{REACToR}

    He Liu, Feng Lu, Alex Forencich, Rishi Kapoor, Malveeka Tewari, Geoffrey M Voelker, George Papen, Alex C Snoeren, and George Porter. Circuit switching under the radar with{REACToR}. In 11th USENIX Symposium on Networked Systems Design and Implementation (NSDI 14), pages 1–15, 2014

    2014

  23. [31]

    Lightwave fabrics: at-scale optical circuit switching for datacenter and machine learning systems

    Hong Liu, Ryohei Urata, Kevin Yasumura, Xiang Zhou, Roy Bannon, Jill Berger, Pedram Dashti, Norm Jouppi, Cedric Lam, Sheng Li, et al. Lightwave fabrics: at-scale optical circuit switching for datacenter and machine learning systems. In Proceedings of the ACM SIGCOMM 2023 Confe...

    2023

  24. [32]

    Quartz: a new design element for low-latency dcns

    Yunpeng James Liu, Peter Xiang Gao, Bernard Wong, and Srinivasan Keshav. Quartz: a new design element for low-latency dcns. ACM SIGCOMM Computer Communication Review , 44(4):283–294, 2014

    2014

  25. [33]

    Expanding across time to deliver band- width efficiency and low latency

    William M Mellette, Rajdeep Das, Yibo Guo, Rob McGuinness, Alex C Snoeren, and George Porter. Expanding across time to deliver band- width efficiency and low latency. In 17th USENIX Symposium on Networked Systems Design and Implementation (NSDI 20) , pages 1–18, 2020

    2020

  26. [34]

    Realizing rotornet: Toward practical microsecond scale optical networking

    William M Mellette, Alex Forencich, Rukshani Athapathu, Alex C Snoeren, George Papen, and George Porter. Realizing rotornet: Toward practical microsecond scale optical networking. In Proceedings of the ACM SIGCOMM 2024 Conference, pages 392–414, 2024

    2024

  27. [35]

    Rotornet: A scalable, low-complexity, optical datacenter network

    William M Mellette, Rob McGuinness, Arjun Roy, Alex Forencich, George Papen, Alex C Snoeren, and George Porter. Rotornet: A scalable, low-complexity, optical datacenter network. In Proceedings of the Conference of the ACM Special Interest Group on Data Communication , pages 26...

    2017

  28. [36]

    Homa: A receiver-driven low-latency transport protocol using network priorities

    Behnam Montazeri, Yilong Li, Mohammad Alizadeh, and John Ouster- hout. Homa: A receiver-driven low-latency transport protocol using network priorities. In Proceedings of the 2018 Conference of the ACM Special Interest Group on Data Communication , pages 221–235, 2018

    2018

  29. [37]

    Mukerjee, Christopher Canel, Weiyang Wang, Daehyeok Kim, Srinivasan Seshan, and Alex C

    Matthew K. Mukerjee, Christopher Canel, Weiyang Wang, Daehyeok Kim, Srinivasan Seshan, and Alex C. Snoeren. Adapting TCP for Reconfigurable Datacenter Networks. In 17th USENIX Symposium on Networked Systems Design and Implementation (NSDI 20) , pages 651–666, Santa Clara, CA, ...

    2020

  30. [38]

    Integrating microsecond circuit switching into the data center

    George Porter, Richard Strong, Nathan Farrington, Alex Forencich, Pang Chen-Sun, Tajana Rosing, Yeshaiahu Fainman, George Papen, and Amin Vahdat. Integrating microsecond circuit switching into the data center. ACM SIGCOMM Computer Communication Review , 43(4):447–458, 2013

    2013

  31. [39]

    Jupiter evolving: transforming google’s datacenter network via optical circuit switches and software-defined networking

    Leon Poutievski, Omid Mashayekhi, Joon Ong, Arjun Singh, Mukarram Tariq, Rui Wang, Jianan Zhang, Virginia Beauregard, Patrick Conner, Steve Gribble, et al. Jupiter evolving: transforming google’s datacenter network via optical circuit switches and software-defined networking. ...

    2022

  32. [41]

    Inside the social network’s (datacenter) network

    Arjun Roy, Hongyi Zeng, Jasmeet Bagga, George Porter, and Alex C Snoeren. Inside the social network’s (datacenter) network. In Pro- ceedings of the 2015 ACM Conference on Special Interest Group on Data Communication, pages 123–137, 2015

    2015

  33. [42]

    Semi-oblivious reconfigurable datacenter networks

    Nitika Saran, Daniel Amir, Tegan Wilson, Robert Kleinberg, Vishal Shrivastav, and Hakim Weatherspoon. Semi-oblivious reconfigurable datacenter networks. In Proceedings of the 23rd ACM Workshop on Hot Topics in Networks, pages 150–158, 2024

    2024

  34. [43]

    Programmable calendar queues for high-speed packet scheduling

    Naveen Kr Sharma, Chenxingyu Zhao, Ming Liu, Pravein G Kannan, Changhoon Kim, Arvind Krishnamurthy, and Anirudh Sivaraman. Programmable calendar queues for high-speed packet scheduling. In Proceedings of NSDI, 2020

    2020

  35. [44]

    c- through: Part-time optics in data centers

    Guohui Wang, David G Andersen, Michael Kaminsky, Konstantina Papagiannaki, TS Eugene Ng, Michael Kozuch, and Michael Ryan. c- through: Part-time optics in data centers. In Proceedings of the ACM SIGCOMM 2010 Conference, pages 327–338, 2010

    2010

  36. [45]

    Topoopt: Optimizing the network topology for distributed dnn training

    Weiyang Wang, Moein Khazraee, Zhizhen Zhong, Zhijao Jia, Dhee- vatsa Mudigere, Ying Zhang, Anthony Kewitsch, and Manya Ghobadi. Topoopt: Optimizing the network topology for distributed dnn training. arXiv preprint arXiv:2202.00433, 2022

    2022 arXiv

  37. [46]

    Masking failures from appli- cation performance in data center networks with shareable backup

    Dingming Wu, Yiting Xia, Xiaoye Steven Sun, Xin Sunny Huang, Sim- barashe Dzinamarira, and TS Eugene Ng. Masking failures from appli- cation performance in data center networks with shareable backup. In Proceedings of the 2018 Conference of the ACM Special Interest Group on Da...

    2018

  38. [47]

    En- abling topological flexibility for data centers using omniswitch

    Yiting Xia, Mike Schlansker, TS Eugene Ng, and Jean Tourrilhes. En- abling topological flexibility for data centers using omniswitch. In HotCloud, 2015

    2015

  39. [48]

    A tale of two topologies: Exploring convertible data center network architectures with flat-tree

    Yiting Xia, Xiaoye Steven Sun, Simbarashe Dzinamarira, Dingming Wu, Xin Sunny Huang, and TS Eugene Ng. A tale of two topologies: Exploring convertible data center network architectures with flat-tree. In Proceedings of the Conference of the ACM Special Interest Group on Data C...

    2017

  40. [49]

    Exrec: Experimental framework for reconfigurable networks based on off-the-shelf hardware

    Johannes Zerwas, Chen Avin, Stefan Schmid, and Andreas Blenk. Exrec: Experimental framework for reconfigurable networks based on off-the-shelf hardware. In Proceedings of the Symposium on Ar- chitectures for Networking and Communications Systems , pages 66–72, 2021. 14 A QUEUE...

    2021

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.