[Paper] BouquetFL: Emulating diverse participant hardware in Federated Learning

Source: arXiv - 2602.06498v1

Overview

Federated Learning (FL) promises privacy‑preserving model training across many devices, yet most research still runs on a single, homogeneous machine. BouquetFL bridges this gap by letting researchers emulate a wide spectrum of client hardware—from low‑end phones to modest lab workstations—on just one physical host. This makes it possible to study how real‑world device diversity impacts FL without the logistical nightmare of gathering dozens of physical devices.

Key Contributions

• Hardware‑aware FL simulation framework that enforces CPU, memory, and GPU constraints to mimic heterogeneous client devices.
• Extensive device profile library covering common consumer gadgets and small‑lab equipment, plus a data‑driven “hardware sampler” that reflects real‑world device popularity.
• Simple, programmatic API for configuring federations, swapping device mixes, and reproducing experiments reproducibly.
• Open‑source implementation that integrates with popular FL toolkits (e.g., TensorFlow Federated, PySyft), lowering the entry barrier for heterogeneity studies.

Methodology

BouquetFL runs all FL participants as isolated processes on a single host. Before each client’s training step, the framework applies Linux cgroups (or Windows Job Objects) to throttle resources—CPU cores, RAM, and GPU memory—according to a chosen device profile. A “hardware sampler” draws random profiles based on publicly available market share statistics, enabling realistic federation compositions. Researchers can also define custom profiles (e.g., a Raspberry Pi‑class node) and plug them into the simulation pipeline with a few lines of Python code. The rest of the FL workflow (model aggregation, communication simulation) remains unchanged, so existing FL codebases can be swapped in with minimal refactoring.

Results & Findings

• Training time variance: Simulated heterogeneous federations exhibited up to 3× slower convergence compared to homogeneous setups, highlighting the bottleneck introduced by low‑end clients.
• Straggler mitigation: Simple client‑selection heuristics (e.g., “skip the slowest 20 %”) recovered up to 85 % of the performance loss without sacrificing model accuracy.
• Resource‑aware scheduling: Dynamically assigning more data to faster profiles (while keeping overall data distribution balanced) yielded a 12 % reduction in wall‑clock training time.
• Reproducibility: Across three independent runs with identical hardware samplers, the variance in final model accuracy stayed within ±0.3 %, confirming the stability of the emulation approach.

Practical Implications

• Product teams can prototype FL deployments on a single workstation, testing how a mix of smartphones, tablets, and edge gateways will affect training latency and energy consumption.
• Framework developers gain a testbed for new straggler‑handling or adaptive aggregation algorithms without needing a physical device lab.
• Ops engineers can benchmark different hardware‑allocation policies (e.g., “favor GPU‑enabled clients”) before committing to costly hardware purchases.
• Regulatory compliance testing becomes easier: teams can simulate worst‑case device capabilities to ensure that privacy guarantees hold even on the most constrained nodes.

Limitations & Future Work

• Network effects are abstracted: BouquetFL focuses on compute heterogeneity; realistic latency, packet loss, and bandwidth constraints are only approximated.
• GPU emulation granularity: Current throttling works at the memory‑allocation level; fine‑grained GPU compute throttling (e.g., CUDA core limits) remains an open challenge.
• Scalability ceiling: While a single host can emulate dozens of clients, scaling to hundreds or thousands may require distributed simulation clusters.
• Future directions include integrating network simulators (e.g., ns‑3), adding power‑consumption models, and extending support for emerging hardware like NPUs and edge‑AI accelerators.

Paper Information

• arXiv ID: 2602.06498v1
• Categories: cs.DC
• Published: February 6, 2026
• PDF: Download PDF