HybridFLow: SDN-Orchestrated Client Partitioning for Hybrid Federated Learning

Osama Abu Hamdan; Rabin Pandey; Hao Che; Engin Arslan; Md Arifuzzaman

Source-linked AI summary

HybridFLow: SDN-Orchestrated Client Partitioning for Hybrid Federated Learning

Osama Abu Hamdan, Rabin Pandey, Hao Che, Engin Arslan, Md Arifuzzaman

arXiv:2609.10404v1Sep 9, 2026cs.NIcs.LG

TL;DR

Wide-area cross-silo FL suffers from network-driven stragglers, while asynchronous participation can impair convergence under non-IID data. HybridFLow uses SDN telemetry and calibrated communication-time estimates to partition clients before each round, then refines those estimates through feedback. Across emulated topologies, it reaches 80% accuracy faster than SmartFLow while reducing round duration, whereas FedAsync does not reach the target.

Problem
Wide-area FL lacks endpoint visibility into shared bottlenecks, path contention, and changing congestion needed for effective hybrid client partitioning.
Method
HybridFLow uses a closed-loop SDN controller to estimate per-client communication times, partition clients by latency and staleness trade-offs, and refine estimates from measured transfers.
Results
HybridFLow reaches the 80% accuracy target 33–40% faster than SmartFLow and reduces average round time by 30–40 seconds across topologies, while FedAsync fails to reach 80% under non-IID data.
Takeaways & Limitations
Network-aware orchestration reduces synchronization blocking while preserving convergence quality under non-IID data distributions in the evaluated emulated cross-silo topologies.
Takeaways & Limitations
The evaluation uses emulated cross-silo environments, and the design assumes topology-wide SDN telemetry, trustworthy measurements, and well-behaved participants.

Abstract

from arXiv · show

Cross-silo Federated Learning (FL) enables geographically distributed institutions to collaboratively train machine learning models without sharing raw data. In wide-area deployments, however, communication delays often dominate round completion time and exacerbate the straggler effect. Hybrid FL addresses this challenge by combining synchronous and asynchronous client participation, but effective partitioning requires visibility into network conditions such as shared bottlenecks, link utilization, and path contention that individual clients cannot observe. We present HybridFLow, a closed-loop SDN-driven orchestration framework that integrates network-layer intelligence directly into hybrid FL. Leveraging the SDN controller's global topology view, HybridFLow generates calibrated per-client communication-time estimates before each training round and uses them to partition clients into synchronous and asynchronous groups while balancing round latency and update staleness. After each round, measured communication times are fed back to the controller to continuously refine future predictions. Experimental results across multiple network topologies show that HybridFLow reaches 80% target accuracy 33-40% faster than SmartFLow and reduces average round duration by 30-40 seconds, while FedAsync fails to reach the target accuracy under non-IID data distributions.

I. INTRODUCTION AND MOTIVATION

Wide-area cross-silo FL is constrained by network heterogeneity: synchronous rounds wait for stragglers, while fully asynchronous aggregation can harm convergence under non-IID data. HybridFLow addresses this gap by using SDN-wide network visibility to guide hybrid partitioning and feedback-driven orchestration.

Motivation: Over 47% longer round completion than optimized routing illustrates how heterogeneous WAN paths amplify synchronous FL’s straggler effect.Synchronous aggregation waits for every participating client, so the slowest path gates throughput.
Motivation: Asynchronous FL removes synchronization barriers but can fail to reach target accuracy under non-IID data because stale and dominant updates degrade convergence.Hybrid FL retains a synchronous subset while allowing slower clients to contribute asynchronously.
Problem: Endpoint measurements cannot reveal shared bottlenecks, cross-client path interference, or evolving background congestion that determine network-induced stragglers.These conditions make sync/async partitioning fundamentally dependent on network state, not only client compute characteristics.
Problem: SDN contributes topology-wide visibility into link utilization, latency, loss, and path contention, enabling network conditions to inform sync/async partitioning.This control-plane view captures shared bottlenecks and interference unavailable to individual FL endpoints.
Contribution: HybridFLow combines calibrated per-client communication-time estimates with cost-minimizing partitioning that balances synchronous latency against asynchronous staleness risk.Measured transfer times refine future estimates after each round through a closed-loop, round-gated mechanism.
Contribution: Unlike model- or update-level communication methods and prior SDN-assisted FL work, HybridFLow uses network telemetry directly for hybrid training-mode decisions.The claimed novelty is topology-wide telemetry as a control signal for cross-silo sync/async client partitioning.

III. SYSTEM OVERVIEW

HybridFLow separates SDN network intelligence from FL orchestration, connecting them through a lightweight interface that supplies per-client path and transfer-time estimates for round-level partitioning. Its scheduler scores shared paths using bandwidth, latency, and loss, while feedback corrects estimates over successive rounds.

Architecture: HybridFLow uses independent SDN-controller and FL layers connected by ZeroMQ, separating network monitoring and estimation from partitioning and aggregation.The SDN layer runs on ONOS, while the FL layer uses Flower; neither accesses the other’s internal state.
Architecture: One estimate-request and measurement-report exchange per round connects network estimation with hybrid aggregation and closes the orchestration feedback loop.The SDN controller supplies estimates before the round, and measured transfer times return afterward.
Path Assignment: The Flow Scheduler assigns one S2C and one C2S path per client while accounting for competing paths, shared links, and changing network conditions.A greedy implementation sorts clients by best available adjusted score and fixes assignments at round start.
Path Assignment: Candidate paths are evaluated using projected fair share, effective RTT, and end-to-end packet loss to estimate comparative network quality.Projected fair share reflects existing flows and capacity consumed by earlier assignments in the same scheduling sweep.
Path Assignment: Effective RTT sums forward and reverse-link latencies because data and acknowledgments traverse opposite directions of the path.The reverse link is denoted by l̄ in the path model.
Path Assignment: Dividing model size D by EffScore(P) produces a raw transfer-time estimate, with ε preventing division by zero when path loss is negligible.EffScore(P) is used as a comparative path-ranking proxy rather than a complete steady-state TCP model.

2) Correction Factors and Adjusted Scoring:

HybridFLow corrects raw network scores with per-client, direction, and path-specific EWMA factors, producing communication-time estimates that improve as observations accumulate.

Correction factors: EWMA correction factors compensate for systematic bias in raw effective scores for each client, direction, and path.The system falls back to aggregate correction data or 1.0 when history is unavailable.
Adjusted scoring: Equation (9) yields calibrated communication-time estimates in seconds that improve as the correction factor converges.

B. FL Layer

The FL layer combines SDN-provided communication estimates with EWMA-smoothed computation times to estimate each client’s total completion time before a round.

FL Layer: The FL layer uses per-client communication-time estimates from the SDN layer to partition clients, orchestrate training, and aggregate updates.
Completion-time estimation: Each client’s estimated completion time combines S2C communication, C2S communication, and EWMA-smoothed local computation time.The computation component is based on observed local computation times.

2) Gradient Importance Tracking:

HybridFLow tracks gradient importance and uses completion-time estimates, normalized latency, and staleness penalties to select a contiguous synchronous prefix through a greedy sweep.

Gradient Importance Tracking: Gradient importance is tracked from client update magnitudes and smoothed globally with a parameter controlling adaptation to recent gradients.The tracker considers clients contributing either synchronous or asynchronous updates.
Gradient Importance Tracking: Clients with importance weights greater than one incur higher asynchronous staleness penalties, favoring synchronous assignment.
Hybrid partitioning: The partitioner minimizes a convex combination of normalized synchronous latency and asynchronous staleness penalty.λ = 0 emphasizes fastest synchronous rounds, λ = 1 emphasizes lower asynchronous penalty, and λ = 0.5 weights both equally.
Hybrid partitioning: All clients are sorted by estimated completion time, allowing the optimal split to be found by a greedy sweep in O(n log n) time.
Hybrid execution: The hybrid server dispatches asynchronous clients immediately, waits for synchronous clients, and buffers asynchronous updates for later aggregation.

4) Hybrid Aggregation:

HybridFLow aggregates synchronous updates first, then integrates buffered asynchronous updates using staleness-aware weighting and a data- and staleness-dependent mixing coefficient.

Hybrid Aggregation: Synchronous results are aggregated with a configured strategy such as FedAvg or FedProx, independently of partitioning and scheduling.
Hybrid Aggregation: Buffered asynchronous updates are integrated using the staleness-aware framework introduced by FedAsync.
Hybrid Aggregation: Each asynchronous update receives a mixing coefficient combining a base rate with configurable polynomial, exponential, or constant staleness decay.The default is polynomial decay with a = 0.5.
Hybrid Aggregation: The asynchronous average is merged with the synchronous model using a coefficient that separates asynchronous data fraction from mean staleness quality.The product remains within [0, 1] without clamping heuristics.

5) Async Dispatch and Re-Dispatch:

Asynchronous clients are dispatched with one queued task at a time, and completed updates trigger timing-aware re-dispatch decisions.

Each asynchronous client has a queue depth of one, so the dispatcher skips it while another task is in flight.Completed updates are placed in a buffer for subsequent processing.
A client finishing within its dispatch round is not redispatched because it may join the synchronous group in the next partition.

C. FL-SDN Data Exchange Protocol

HybridFLow exchanges network estimates before each round and reports measured communication times afterward, coordinating SDN scheduling with FL partitioning through a round-gated protocol.

The protocol exchanges two message types once per round via ZeroMQ before any data transfer begins.
At round start, the controller returns adjusted S2C and C2S communication-time estimates for every available client.
The server computes each client’s completion time as estimated S2C time plus estimated C2S time plus average computation time.
The server partitions clients into synchronous and asynchronous groups using completion times, importance weights, and configured partitioning parameters.
After each round, measured communication times are sent to the controller to update EWMA correction factors and improve future estimates.

D. End-to-End Round Execution

Each HybridFLow round estimates communication times, partitions clients before training, aggregates synchronous and asynchronous updates, and reports measurements for subsequent calibration.

Algorithm 2 integrates the SDN and FL layers through the per-round data exchange protocol.
Path assignment and partitioning finish before training, preventing mid-round switching and its associated measurement instability.
The Hybrid Partitioner computes total completion times and greedily minimizes J(S) over clients sorted by completion time.
Asynchronous clients train concurrently in the background while the server waits for all synchronous clients to complete.
Two-stage aggregation first produces a synchronous model, then merges buffered asynchronous updates using staleness-weighted averaging.
Measured communication times update EWMA correction factors, improving path-score calibration in later rounds.

IV. EXPERIMENTAL RESULTS

Experiments evaluate HybridFLow across three emulated, increasingly large topologies under controlled non-IID workloads, comparing network-aware hybrid partitioning with synchronous and asynchronous baselines.

Three topologies scale from E1 with 15 switches and clients to E3 with 35 switches and clients under time-varying background congestion.
CIFAR-10 uses a pathological non-IID partition assigning seven of ten classes to each client, with MobileNetV3-Large generating 14.2 MB communication workloads.
30–40 seconds: HybridFLow reduces average round time relative to SmartFLow in every topology.
SmartFLow is fully synchronous, FedAsync is fully asynchronous, and HybridFLow uses network-aware hybrid partitioning with shared aggregation settings.
Training targets 80% server evaluation accuracy, while methods missing the target within budget are reported at final achieved accuracy.

B. Results

HybridFLow reduces synchronization blocking through network-aware, continuously calibrated client partitioning. Across evaluated topologies, it shortens round duration and reaches target accuracy faster than SmartFLow while FedAsync fails under non-IID data.

Synchronization reduction: HybridFLow reclassifies straggling clients as asynchronous, closing rounds without discarding their updates and reducing synchronization delay rather than communication volume.Asynchronous updates are incorporated in subsequent rounds with a staleness penalty.
Round duration: 30–40 seconds: average round time drops by this amount relative to SmartFLow in every topology.HybridFLow has narrower round-duration variation in E1 and E2, while E3 retains greater variability at 35-client scale.
Adaptive partitioning: The partitioner continuously adjusts the synchronous boundary in response to topology-dependent network conditions rather than applying a static policy.Async fractions vary with network heterogeneity across E1, E2, and E3.
Prediction accuracy: Communication MAE is 7.2 s in E1, 6.2 s in E2, and 9.2 s in E3, while computing MAE stays below 1.5 s across all topologies.EWMA correction calibrates future SDN estimates from measurements reported after each round.
Accuracy and convergence: 33–40% faster: HybridFLow reaches the 80% accuracy target than SmartFLow across all three topologies, whereas FedAsync never reaches it under non-IID data.The results link reduced synchronization blocking with preserved convergence quality.
Limitations: The evaluation remains limited to an emulated cross-silo environment, with validation at larger scales and across operational WANs left for future work.The framework also assumes topology-wide SDN telemetry and trustworthy telemetry and participants.