Flux: Optimal Scheduling of Optical Circuit Switches for LLM Training
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Computer Science. The day in computer science: software, systems and theory on arXiv.
Tom: I'm Tom.
Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.
Tom: We're looking at the paper: "Flux: Optimal Scheduling of Optical Circuit Switches for LLM Training".
Paper summary: Tom: So, looking at the whole thing, the authors of Flux have developed a scheduler that optimally schedules optical circuit switches based on the structure of the entire workload <ref:2609.25949#pg4>. They’ve shown it can reduce training iteration time by up to ten times and peak NIC buffer requirements by more than three orders of magnitude compared to traditional periodic schedulers <ref:2609.25949#pg2>.
Jane: And the key thing here is that Flux achieves this because it schedules circuits together with the workload, allowing communication and reconfiguration to overlap with compute <ref:2609.25949#pg4>. It's not just about optimizing for traffic; it’s about aligning the network actions with the actual computation timeline <ref:2609.25949#pg3>.
Lu: The paper formalizes this by creating a Mixed Integer Linear Program that takes the system of accelerators and a specific workload as inputs, and its objective is to minimize the total completion time of that workload graph, defined as the minimum of the maximum time across all those points and vertices in the graph <ref:2609.25949#pg4>.
Meng: What this means for us is that we need to think about modeling our workloads in a way that captures these temporal dependencies accurately if we want to leverage these types of scheduling techniques <ref:2609.25949#pg7>. The authors are also looking ahead at turning Flux into a tool for scheduling large and realistic problem sizes, which includes designing intelligent heuristics and supporting MoE models <ref:2609.25949#pg4>.
Lalam: I think the paper really proves that considering the temporal structure of the workload is critical for achieving better efficiency in optical circuit switching for LLM training <ref:2609.25949#pg2>. It shows that what matters is whether a reconfiguration enables useful circuits <ref:2609.25949#pg4>.
Tom: So, the authors are showing that the separation between network scheduling and compute can be a fundamental inefficiency, and Flux addresses it by coupling them into one schedule <ref:2609.25949#pg3>. It’s a big step in making high-density optical networking practical for large scale AI training <ref:2609.25949#pg1>.
Jane: The implication is that we move beyond just looking at aggregate traffic to designing schedules that actively pipeline communication with computation, which really unlocks the potential of these systems <ref:2609.25949#pg3>.
Conclusion: Tom: So, Flux is about finding the best way to schedule those optical circuit switches for training large language models <ref:2609.25949#pg4>.
Jane: It’s essentially taking a big task—the whole workload of training—and figuring out how to map it onto the network hardware in a way that keeps everything running smoothly, even when the network has to reconfigure itself <ref:2609.25949#pg2>.
Lu: The authors use a mixed integer linear program to look at the entire training graph, including both how fast things compute and how they communicate <ref:2609.25949#pg4>. They’re minimizing the total time it takes for that whole training process to finish <ref:2609.25949#pg4>.
Meng: It sounds like they’re trying to make sure the network reconfiguration doesn't just sit there and cause delays while the AI accelerators are waiting for data <ref:2609.25949#pg1>. That idle time costs money at scale, so that matters to me <ref:2609.25949#pg4>.
Lalam: What they show is that when you schedule the network and the compute together, you can actually reduce iteration times by up to ten times and shrink those massive memory requirements for the NICs by more than three orders of magnitude <ref:2609.25949#pg2>.
Tom: Ten times faster iterations is a big number, but I want to know what they mean in practice for someone trying to deploy a model right now <ref:2609.25949#pg1>.
Jane: It means instead of waiting around for the network to catch up, the communication happens while the compute part is already working, which cuts down on those bottlenecks <ref:2609.25949#pg4>.
Lu: They’re basically saying that what really matters isn't how often you reconfigure a switch, but whether that reconfiguration lets you use the circuit in a useful way during the compute phase <ref:2609.25949#pg7>.
Meng: If this works well across different hardware, then it means we can actually move toward more efficient optical interconnects for massive AI training jobs <ref:2609.25949#pg4>.
Lalam: I'm confident that the connection between the workload structure and the network schedule is where the real power of this work is, especially when you think about scheduling huge models with many experts <ref:2609.25949#pg7>.
IDLab, University of Antwerp - imec
cs.NI, cs.DC
Submitted: 2026-09-22
Updated: 2026-11-17
Comments: Moved to IEEE format, added author, moved order of sections, small changes
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 92/100
The gist: Optical Circuit Switching (OCS) offers high bandwidth density and energy efficiency for LLM training, but incurs a non-negligible reconfiguration delay <ref:2609.25949#pg2> This paper presents Flux…
Key concepts
- Optical Circuit Switching (OCS)
- A technology that uses light beams to create dedicated, high-bandwidth paths through a network. It offers high bandwidth density for LLM training but introduces a delay when switching circuits, which Flux aims to minimize.
- Workload-Aware Scheduling
- The method of scheduling switches by considering the entire structure of the LLM training job, not just traffic demand. It treats compute and communication as a single problem to find the best overall execution order and circuit configuration.
- Mixed Integer Linear Program (MILP)
- A mathematical framework used by Flux to formally define the scheduling problem. It takes the system's accelerators and the workload as inputs, aiming to minimize total completion time while respecting physical constraints like reconfiguration delays.
- Directed Acyclic Graph (DAG)
- A way to represent the LLM training process where each task is a point-to-point send between accelerators. Edges in this graph encode dependencies, including the necessary compute time, which helps Flux determine the correct execution sequence.
Terminology
Summary
Optical Circuit Switching (OCS) offers high bandwidth density and energy efficiency for LLM training, but incurs a non-negligible reconfiguration delay <ref:2609. p. 025949#pg2> This paper presents Flux, a scheduler that optimally schedules optical circuit switches based on the structure of the entire workload <ref:2609. p. 025949#pg4> Flux reduces training iteration time by up to 10× and peak NIC buffer requirements by more than three orders of magnitude compared to traditional periodic schedulers <ref:2609. p. 025949#pg2>
Critique of Traditional Scheduling Methods
Prior work typically schedules optical circuit switches independently of compute, using aggregate traffic demand to determine which circuits to provision and when <ref:2609.25949#pg2> This separation creates a fundamental inefficiency because reconfigurations that ignore the compute timeline can stall communication, resulting in low circuit utilization and large buffer requirements <ref:2609. p. 025949#pg2> Schedulers relying on traffic matrices cannot capture compute-communication overlap, meaning the network may perform large transfers while accelerators sit idle waiting for data <ref:2609. p. 025949#pg2> Furthermore, separating communication scheduling from compute creates an inefficiency because schedulers that rely on traffic matrices cannot capture compute-communication overlap <ref:2609. p. 025949#pg2>
Formalizing Workload-Aware Scheduling
Flux formalizes workload-aware circuit scheduling as a Mixed Integer Linear Program (MILP) that consolidates compute and communication into a single circuit schedule <ref:2609.25949#pg2> The formulation takes two inputs: a system of n accelerators interconnected through an OCS network and a workload (a specific LLM training iteration) <ref:2609. p. 225949#pg4> The objective function minimizes the total completion time of the workload graph, defined as min max (tv + pv) <ref:2609. p. 225949#pg4> Constraints fix the execution order of every pair of tasks with a direct edge in the graph, independently of which switch they are assigned to <ref:2609. p. 325949#pg4>
The Role of Temporal Structure and Dependencies
The workload model is represented as a Directed Acyclic Graph (DAG) where each vertex is an atomic point-to-point send between two accelerators and each edge encodes a dependency, including compute time <ref:2609.25949#pg4> DAG dependencies fix the execution order of every pair of tasks with a direct edge in the graph, independently of which switch they are assigned to <ref:2609. p. 325949#pg4> The constraints capture physical limitations related to reconfiguration when two tasks share an endpoint and are scheduled on the same switch <ref:2609. p. 325949#pg4>
Performance Evaluation and Results
Flux consistently achieves the lowest iteration time across the evaluated reconfiguration delays, while significantly reducing peak NIC buffer demand by more than three orders of magnitude relative to the other schedulers <ref:2609.25949#pg2> The results show that Flux schedules circuits together with the workload, allowing communication and reconfiguration to be overlapped with compute <ref:2609. p. 425949#pg4> A major improvement of Flux is in peak on-chip NIC memory demand, as it keeps buffer requirements low <ref:2609. p. 425949#pg4>
Conclusion and Future Directions
Flux demonstrates the effectiveness of workload-aware circuit scheduling for LLM training by considering the temporal structure of the workload <ref:2609.25949#pg2> In future work, the objective will be to turn Flux from a powerful formalization into a tool that can be used to schedule large and realistic problem sizes <ref:2609. p. 525949#pg4> This encompasses the design of intelligent heuristics, support for MoE models, and an extensive evaluation campaign <ref:2609. p. 525949#pg4>
Related Work Context
Existing circuit scheduling approaches typically rely on deriving a schedule from an aggregate traffic demand matrix, which discards the ordering of communication operations <ref:2609.25949#pg2> Flux schedules circuit reconfigurations and communication while explicitly taking into account when compute occurs <ref:2609. p. 525949#pg4> The paper highlights that what matters is not how often the network reconfigures, but whether a reconfiguration enables useful circuits <ref:2609. p. 525949#pg4> The authors observe that Flux can be effectively used across a wide variety of OCS switching technologies <ref:2609. p. 525949#pg4>
References
[1] Vamsi Addanki, Chen Avin, and Stefan Schmid. 2023. Mars: NearOptimal Throughput with Shallow Buffers in Reconfigurable Datacenter Networks. p. 5 Proc. ACM Meas. Anal. Comput. Syst. 7, 1 (March 2023), 2:1–2:43 <ref:2609. p. 525949#pg8>
[2] Hitesh Ballani, Paolo Costa, Raphael Behrendt, Daniel Cletheroe, Istvan Haller, Krzysztof Jozwik, Fotini Karinou, Sophie Lange, Kai 6. 2020. Sirius: A Flat Datacenter Network with Nanosecond Optical Switching. p. 6 In Proceedings of the ACM SIGCOMM 2020 Conference (SIGCOMM ’20). p. 6 Association for Computing Machinery, New York, NY, USA, 782–797 <ref:2609. p. 625949#pg8>
[3] Mokhtar S. Bazaraa, John J. p. 6 Jarvis, and J. p. 6 J. Jarvis. 1977. Linear Programming and Network Flows. p. 6 Wiley, New York <ref:2609. p. 625949#pg4>
[4] Nathan Farrington, George Porter, Sivasankar Radhakrishnan, Hamid Hajabdolali Bazzaz, Vikram Subramanya, Yeshaiahu Fainman, George Papen, and Amin Vahdat. 2010. Helios: A Hybrid Electrical/Optical Switch Architecture for Modular Data Centers. p. 6 In Proceedings of the ACM SIGCOMM 2010 Conference (SIGCOMM ’10). p. 6 Association for Computing Machinery, New York, NY, USA, 339–350 <ref:2609. p. 625949#pg5>
[5] William Fedus, Barret Zoph, and Noam Shazeer. 2022. Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity. p. 6 The Journal of Machine Learning Research 23, 1 (Jan. p. 6 2022), 120:5232–120:5270 https://dl. p. 6acm.org/doi/10. p. 6 p. 65555/3586589 <ref:2609. p. 625949#pg7>
[6] Adithya Gangidi, Rui Miao, Shengbao Zheng, Sai Jayesh Bondu, Guilherme Goes, Hany Morsy, Rohit Puri, Mohammad Riftadi, Ashmitha Jeevaraj Shetty, Jingyi Yang, Shuqiang Zhang, Mikel Jimenez Fernandezhadernege 2024. RDMA over Ethernet for Distributed Training at Meta Scale. p. 6 In Proceedings of the ACM SIGCOMM 2024 Conference. p. 6 p. 6 p. 6 ACM, Sydney NSW Australia, 57–70 doi: 10. p. 61145/3651890. p. 6 p. 6 p. 63672233 <ref:2609. p. 025949#pg7>
[7] Garrett Birkhoff. 1946. Tres Observaciones Sobre El Algebra Lineal. p. 6 Univ. Nac. Tucuman, Ser. p. 6 A 5 (1946), 147–154 <ref:2609. p. 625949#pg3>
[8] Amir Gholami, Zhewei Yao, Sehoon Kim, Coleman Hooper, Michael W. Mahoney, and Kurt Keutzer. p. 6 2024. AI and Memory Wall. p. 6 IEEE Micro 44, 3 (May 2024), 33–39 doi: 10. p. 61109/MM.2024.3373763
[9] Aaron Grattafiori et al. 2024. The Llama 3 Herd of Models. arXiv:2407. p. 621783 [cs]
[10] Wenzhe Guo, Joyjit Kundu, Uras Tos, Weijiang Kong, Giuliano Sisto, Timon Evenblij, and Manu Perumkunnil. 2025. Keeping up with Large Language Models: A Holistic Methodology of Compute, Memory, Communication, and Cost Modeling. p. 6 In 2025 IEEE International Symposium on Workload Characterization (IISWC). p. 6 116–126 doi: 10. p. 61109/IISWC66894. p. 62025.00019 <ref:2609. p. 025949#pg8>
[11] Gurobi Optimization, LLC. 2026. Gurobi Optimizer. p. 6 p. 6 https://www. p. 6gurobi.com Version 13. p. 4 p. 40.2 <ref:2609. p. 025949#pg8>
[12] Norm Jouppi, George Kurian, Sheng Li, Peter Ma, Rahul Nagarajan, Lifeng Nai, Nishant Patil, Suvinay Subramanian, Andy Swing, Brian Towles, Clifford Young, Xiang Zhou <ref:2609.25949#pg8>
[13] Mehrdad Khani et al. 2021. SiP-ML: High-Bandwidth Optical Network Interconnects for Machine Learning Training. p. 6 In Proceedings of the 2021 ACM SIGCOMM 2021 Conference (SIGCOMM ’21). p. 6 Association for Computing Machinery, New York, NY, USA, 657–675 doi: 10. p. 61145/3452296. p. 63472900
[14] Wenxue Li et al. 2024. Understanding Communication Characteristics of Distributed Training. p. 6 In Proceedings of the 8th Asia-Pacific Workshop on Networking (APNet ’24). p. 6 Association for Computing Machinery, New York, NY, USA, 1–8 doi: 10. p. 61145/3663408. p. 63663409 <ref:2609. p. 025949#pg8>
[15] He Liu et al. 2015. Scheduling Techniques for Hybrid Circuit/Packet Networks. p. 6 In Proceedings of the 11th ACM Conference on Emerging Networking Experiments and Technologies (CoNEXT ’15). p. 6 Association for Computing Machinery, New York, NY, USA, 1–3 doi: 10. p. 61145/2716281. p. 62836126 <ref:2609. p. 025949#pg8>
[16] Hong Liu et al. 2023. Lightwave Fabrics: At-Scale Optical Circuit Switching for Datacenter and Machine Learning Systems. p. 6 In Proceedings of the ACM SIGCOMM 2023 Conference. p. 6 ACM, New York NY USA, 499–515 doi: 10. p. 61145/3603269. p. 63604836 <ref:2609. p. 025949#pg8>
[17] William M. Mellette et al. p. 6 p. 6 p. 6 2020. Expanding across Time to Deliver Bandwidth Efficiency and Low Latency. p. 6 In 17th USENIX Symposium on Networked Systems Design and Implementation (NSDI 20). p. 6 1–18 https://www. p. 6usenix.org/conference/nsdi20/presentation/mellette <ref:2609. p. 625949#pg8>
[18] William M. Mellette et al. 2024. Realizing RotorNet: Toward Practical Microsecond Scale Optical Networking. p. 6 In Proceedings of the ACM SIGCOMM 2024 Conference. ACM, Sydney NSW Australia, 392–414 doi: 10. p. 61145/3651890.3672273
[19] William M. Mellette et al. 2017. RotorNet: A Scalable, Low-complexity, Optical Datacenter Network. p. 6 In Proceedings of the Conference of the ACM Special Interest Group on Data Communication. p. 6 ACM, Los Angeles CA USA, 267–280 doi: 10. p. 61145/3098822. p. 63098838 <ref:2609. p. 025949#pg7>
[20] George Porter et al. 2013. Integrating Microsecond Circuit Switching into the Data Center. p. 6 In Proceedings of the ACM SIGCOMM 2013 Conference on SIGCOMM (SIGCOMM ’13). p. 6 Association for Computing Machinery, New York, NY, USA, 447–458 doi: 10. p. 61145/2486001. p. 62486007 <ref:2609. p. 025949#pg8>
[21] Leon Poutievski et al. 2022. Jupiter Evolving: Transforming Google’s Datacenter Network via Optical Circuit Switches and Software-Defined Networking. p. 6 In Proceedings of the ACM SIGCOMM 2022 Conference (SIGCOMM ’22). p. 6 Association for Computing Machinery, New York, NY, USA, 66–85 doi: 10. p. 61145/3544216. p. 63544265 <ref:2609. p. 025949#pg8>
[22] Kun Qian et al. 2024. Alibaba HPN: A Data Center Network for Large Language Model Training. p. 6 In Proceedings of the ACM SIGCOMM 2024 Conference. ACM, Sydney NSW Australia, 691–706 doi: 10. p. 61145/3651890.3672265 <ref:2609. p. 025949#pg8>
[23] Samyam Rajbhandari et al. 2020. ZeRO: Memory Optimizations toward Training Trillion Parameter Models. p. 6 In Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis (SC ’20). p. 6 IEEE Press, Atlanta, Georgia, 1–16 https://dl. p. 6acm.org/doi/10.5555/3433701. p. 63433727
[24] Sundararajan Renganathan and Nick McKeown. 2025. Chronos: Prescheduled Circuit Switching for LLM Training. p. 6 In Proceedings of the 2nd Workshop on Networks for AI Computing (NAIC ’25). p. 7 Association for Computing Machinery, New York, NY, USA, 89–97 doi: 10. p. 61145/3748273. p. 73749210
[25] Noam Shazeer et al. 2017. Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer. p. 7 In International Conference on Learning Representations. p. 7 https://openreview. p. 7net/forum?id=B1ckMDqlg
[26] Mohammad Shoeybi et al. 2020. Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism. p. 7 arXiv:1909. p. 708053 [cs] <ref:2609. p. 025949#pg7>
[27] Srinivas Sridharan et al. 2023. Chakra: Advancing Performance Benchmarking and Co-design Using Standardized Execution Traces. p. 7 arXiv:2305. p. 714516 [cs.LG]
[28] Ryohei Urata et al. 2022. Mission Apollo: Landing Optical Circuit Switching at Datacenter Scale. p. 7 arXiv:2208. p. 710041 [cs.NI] <ref:2609. p. 025949#pg7>
[30] Weiyang Wang et al. 2023. TopoOpt: Co-optimizing Network Topology and Parallelization Strategy for Distributed Training Jobs. p. 7 In 20th USENIX Symposium on Networked Systems Design and Implementation (NSDI 23). p. 7 739–767 https://www. p. 7usenix.org/conference/nsdi23/presentation/wangweiyang
[31] Yuyang Wang et al. [n. d.]. Co-Designed Silicon Photonics Chip I/O for Energy-Efficient Petascale Connectivity. p. 7 15, 8 ([n. d.]), 1581–1591 doi: 10. p. 71109/TCPMT. p. 72024.3492189 <ref:2609. p. 025949#pg7>
[32] William Won et al. 2023. ASTRA-sim2. p. 70: Modeling Hierarchical Networks and Disaggregated Systems for Large-model Training at Scale. p. 7 In 2023 IEEE International Symposium on Performance Analysis of Systems and Software (ISPASS). p. 7 283–294 doi: 10. p. 61109/ISPASS57527. p. 72023.00035 <ref:2609. p. 025949#pg8>
[33] Zhenguo Wu et al. 2025. ACTINA: Adapting Circuit-Switching Techniques for AI Networking Architectures. p. 7 In Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis. p. 6 ACM, St. Louis MO USA, 1211–1222 doi: 10. p. 71145/3712285. p. 73759842
[34] Gurobi Optimization, LLC. 2026. Gurobi Optimizer. Version 13.0.2 https://www. p. 6gurobi.com <ref:2609. p. 0
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Workload-aware circuit scheduling using Flux allows for
compute-communication overlap, and amortization of the reconfiguration delay.
This enables communication operations to occur while compute is running, which reduces training iteration time by up to 10 times compared to traditional periodic schedulers. -
The system can significantly reduce NIC buffer requirements by more than three orders of magnitude compared to traditional periodic schedulers because Flux
schedules circuits based on the entire workload structure, keeping buffer requirements low.
-
The scheduler can be effective across a wide range of switching speeds by
reusing circuits and amortizing reconfiguration delay behind compute and communication.
This ensures that reconfigurations are hidden behind compute phases rather than stalling communication. -
The system can produce schedules that
naturally lead to circuit reuse
by deriving the schedule from the workload graph structure, avoiding the inefficiency where demand-aware approaches cannot capturecompute-communication overlap.
-
The MILP formulation provides a formal framework to minimize
the total completion time of the workload graph: min max (tv + pv).
This allows for optimal assignment of tasks to switches while respecting all DAG dependencies and reconfiguration penalties.
Sources
- The Llama 3 Herd of Models
- Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism
- Chakra: Advancing Performance Benchmarking and Co-design using Standardized Execution Traces
- Mission Apollo: Landing Optical Circuit Switching at Datacenter Scale
Related papers
- Predicting Networks Before They Happen: Experimentation on a Real-Time V2X Digital Twin
- MultiMoQ: Multi-Access Media-Over-QUIC for Robust Immersive Video Streaming
- HiFiNet: Hierarchical Fault Identification in Wireless Sensor Networks via Edge-Based Classification and Graph Aggregation
- Tutorial on Closed-Loop and Application-Aware Digital Twins for Future Wireless Networks
- Claude Code Complete User Handbook
- Bringing dApps to OCUDU: An E3 Controller for Real-Time Open RAN Intelligence