Efficiently Distributed Federated Learning
Gianluca Mittone, Robert Birke, Marco Aldinucci
cs.PF · cs.AI · cs.DC
Quick verdict
Vision paper for C/C++ FL framework FFL claiming 2.5–3.69x speedups over OpenFL on x86/ARM/RISC-V; promising but evaluation is thin and mostly future work.
Summary
This Euro-Par 2023 workshop paper (6 pages) presents FastFederatedLearning (FFL), an open-source federated learning (FL) framework written in C/C++ on top of the FastFlow parallel framework, Cereal serialization, and the PyTorch C++ API. The authors argue that prevailing FL frameworks (FedML, NVIDIA FLARE, Intel OpenFL, Flower, FATE) hard-code the federation communication graph, largely support only DNNs, and are written in plain Python, sacrificing distributed-performance headroom. FFL's headline result is a benchmark on a real heterogeneous cluster spanning x86-64, ARM-v8, and RISC-V devices, where it reports 2.5x–3.69x speedups over Intel OpenFL. A companion effort, OpenFL-extended, ports the model-agnostic AdaBoost.F algorithm into OpenFL and achieves a 5.5x speedup over stock OpenFL after tuning buffers, serialization, and internal structures. However, the paper is primarily a vision/roadmap: the Python Keras-style interface, RISC-pb2l-based middleware with a symbolic reasoner, proxy nodes, dynamic communication graphs, FLaaS, and asynchronous FL are all future work, and the benchmark is reported without dataset, model, topology, client counts, hardware specifics, repetitions, or accuracy/convergence comparison — making the speedup claims impossible to assess from the text. As a position paper it usefully frames gaps in FL tooling, but as evidence of superiority it remains preliminary.
Review complete
Assessment at a glance
4
risk flags identified
● high 1
● medium 2
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 18 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Current FL frameworks hard-code the federation communication graph (usually one fixed topology), preventing users from specifying custom or dynamic federations (§1).
- FFL, a C/C++ framework built on FastFlow, Cereal, and the PyTorch C++ API, achieves 2.5x-3.69x speedups over Intel OpenFL when training on a real heterogeneous cluster of x86-64, ARM-v8, and RISC-V devices.
- OpenFL-extended is claimed to be the first model-agnostic FL framework, supporting both DNNs and traditional ML models via the AdaBoost.F algorithm, optimized to run 5.5x faster than stock OpenFL.
- Python-implemented FL frameworks introduce avoidable delays, and FL frameworks should follow the ML-library pattern of a high-performance C/C++ core behind a friendly Python interface (§1, §2).
Strengths
Validation on genuinely heterogeneous hardware
Section 2 reports FL training actually run across x86-64, ARM-v8 and RISC-V devices on a real cluster — the paper itself notes (§2) that most FL work only simulates on small homogeneous clusters, and it escapes that trap for its headline experiment.
Open-source artifacts
Both FFL and OpenFL-extended are released with public GitHub repositories (footnotes 1-2, github.com/alpha-unito/...), enabling external inspection of the claimed implementations.
Sharp, accurate framing of FL framework deficits
Roadmap grounded in cited literature
Candid self-assessment
Weaknesses
Benchmark reported with zero methodology
No model-quality or convergence comparison
Baseline-fairness and ablation risk
Most claimed capabilities are future work
OpenFL-x 5.5x gain is anecdotal
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This is a 6-page Euro-Par 2023 workshop paper (Springer LNCS, DOI 10.1007/978-3-031-48803-0_40) from the University of Turin group — best classified as a position/vision paper carrying one preliminary benchmark. It advances four claims: (1) existing FL frameworks hard-code the federation communication graph (§1 cites FedML, NVIDIA FLARE, OpenFL, Flower, and HP Swarm Learning as mutually non-interchangeable); (2) most support only DNNs, with GBDT/XGBoost exceptions in FATE and FLARE; (3) Python-implemented frameworks waste distributed-performance headroom through 'poor distributed programming'; (4) FFL, their C/C++ framework on FastFlow + Cereal + the PyTorch C++ API, beats Intel OpenFL by 2.5x-3.69x on a heterogeneous x86-64/ARM-v8/RISC-V cluster. A companion artifact, OpenFL-extended, is claimed as 'the first FL framework to be model-agnostic' via AdaBoost.F, optimized 5.5x over stock OpenFL. Most of §2 is explicitly future work, and the conclusions describe FFL as 'in-development research software'.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
How much faster is FastFederatedLearning (FFL) than Intel OpenFL?
The paper reports FFL training 2.5x to 3.69x faster than Intel OpenFL on a real heterogeneous cluster combining x86-64, ARM-v8, and RISC-V devices. However, the paper gives no dataset, model, topology, or hardware details for the benchmark and no accuracy comparison, so the figures are indicative rather than independently verifiable.
Is the FFL federated learning framework's code available?
Yes. The paper describes FFL as open source at github.com/alpha-unito/FastFederatedLearning, and the companion OpenFL-extended project at github.com/alpha-unito/OpenFL-extended. Both are described as research software still under development.
Is FFL ready for production use?
No. The authors explicitly call FFL 'in-development research software'. The planned features — a Keras-style Python interface, dynamic communication graphs, support for traditional (non-deep) ML models, asynchronous communication, and middleware for multiple backends — are all listed as future work in the paper.
What is OpenFL-extended and what is AdaBoost.F?
OpenFL-extended is a Python extension the authors built that adds the model-agnostic AdaBoost.F federated algorithm to Intel's OpenFL framework; AdaBoost.F extends AdaBoost so any traditional ML model can serve as a weak learner in a federation. After tuning buffers, serialization, and internal data structures, the authors report a 5.5x speedup over the base OpenFL.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- No figures or tables appear in the extracted text; if the published Springer version contains evaluation tables, they were not analyzable here.
- This is the accepted version of a 6-page Euro-Par 2023 workshop paper (LNCS), so brevity and thin evaluation are partly inherent to the venue format.
- Extraction interleaved page headers with body text, but content is complete through the reference list.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.