Home / Curated / Algorithms / FedBuff (Federated Learning with Buffered Asynchronous Aggregation)

FedBuff (Federated Learning with Buffered Asynchronous Aggregation)

Algorithms AIMech:FedBuffBufferedAsynchronousFederatedAggregation DRAFT

Aliases: FedBuff, Buffered Asynchronous Federated Aggregation, Federated Learning with Buffered Asynchronous Aggregation

A buffered asynchronous federated optimization algorithm (Nguyen et al. 2022, Section 3, Algorithm 1) in which the server schedules up to M concurrent clients asynchronously without barrier synchronization, accumulates each arriving client's local multi-step update Delta_i = w^{(t - tau_i)} - w_{i, Q}^{(t - tau_i)} inside a secure aggregation buffer u = u + s(tau_i) Delta_i of configurable capacity K >= 1 (independent of concurrency M), and triggers a global server update w^{(t+1)} = w^{(t)} - eta_g (u / K) only when K updates have been buffered.

Identity Notes & Concept Scope Distinct from synchronous Federated Averaging / FedProx (AIMech:FedProxProximalHeterogeneousFederatedOptimization) and FedOpt (AIMech:FedOptAdaptiveServerOptimizationFedAdagradFedYogiFedAdam), which block each round until the slowest scheduled straggler finishes, and from unbuffered asynchronous FL (FedAsync, K = 1), which updates the global model on every single client arrival and exposes each individual client update without cohort secure aggregation.

Claims & Grounded Evidence

4 Claims

Mechanism #r419-k-fold-server-step-staleness-contraction-and-nonconvex-ergodic-bound

Buffering K >= 1 client updates before advancing the server step index t divides the effective server-step staleness tau_{i, K} by a factor of K relative to unbuffered asynchronous aggregation (tau_{max, K} <= ceil(tau_{max, 1} / K), Appendix A), and under bounded gradient dissimilarity and staleness establishes the smooth non-convex convergence bound O(1 / sqrt(T Q K)) + O(Q^2 tau_max^2 / T) (Section 4, Theorem 1, Corollary 1).

supports Source: Federated Learning with Buffered Asynchronous Aggregation · Locator: Section 4 Convergence Analysis (Assumptions 1-4, Theorem 1, Corollary 1) & Appendix A-B (Relationship Between K and Staleness, Proof of Theorem 1)

Mechanism #r419-secure-aggregation-compatibility-and-batch-learning-rate-normalization

Accumulating K updates inside a Trusted Execution Environment or cryptographic Secure Aggregation buffer hides individual client updates from the server until K clients are averaged (unlike FedAsync with K = 1), while Learning Rate Normalization eta_q = eta_l (b_q / B) equalizes local SGD progress across clients with heterogeneous local batch sizes b_q (Section 3.1 & Appendix C.6.1).

supports Source: Federated Learning with Buffered Asynchronous Aggregation · Locator: Section 3.1 Privacy and Security (Table 1) & Appendix C.6.1 Learning Rate Normalization

Limitation #r419-buffer-size-latency-vs-staleness-tradeoff-and-over-concurrency-limitation

Choosing K too large increases wall-clock waiting time per server update (approaching synchronous FedAvg when K = M), whereas setting concurrency M >> K inflates worst-case staleness tau_max and requires staleness down-weighting s(tau) = (1 + tau)^{-alpha} to prevent stale client updates from destabilizing training.

supports Source: Federated Learning with Buffered Asynchronous Aggregation · Locator: Section 4-5 Trade-offs of Buffer Size K and Concurrency M (Figures 4-6)

Application #r419-celeba-sent140-cifar10-and-production-federated-application

Applied to asynchronous federated learning on CelebA, Sent140, CIFAR-10, and million-device production mobile benchmarks under realistic client speed and availability traces, achieving up to 3.8x speedup over synchronous FedAvg and FedAdam.

supports Source: Federated Learning with Buffered Asynchronous Aggregation · Locator: Section 5 Experiments (Figures 2-6, Tables 2-3), pp. 3586-3590

⚠️ Curated Limitations & Negative Results

Convergence guarantee depends on a bounded maximum staleness tau_max across concurrent asynchronous clients.
Requires tuning the buffer capacity K relative to system concurrency M to balance update frequency against staleness.

Artificial Intelligence Ontology (AIO) Grounding

Mapped AIO Classes

AIO Target Class	Mapping Relation	Mapping Rationale
`AIO:FederatedLearning` Federated Learning	broad_match	FedBuff is a buffered asynchronous aggregation protocol for Federated Learning.

Evidence Sources & Primary Literature

3 Sources

Source ID	Publication / Document	Type / Access	Reference & Checksum
`r419-primary-pdf`	Federated Learning with Buffered Asynchronous Aggregation John Nguyen, Kshitiz Malik, Hongyuan Zhan, Ashkan Yousefpour, Michael Rabbat, Mani Malek, Dzmitry Huba	primary paper full_text	ID: https://proceedings.mlr.press/v151/nguyen22b.html sha256:a979ab1f6bc0…
`r419-arxiv-abs`	Federated Learning with Buffered Asynchronous Aggregation (arXiv abstract) John Nguyen, Kshitiz Malik, Hongyuan Zhan, Ashkan Yousefpour, Michael Rabbat, Mani Malek, Dzmitry Huba	primary paper abstract_only	ID: https://arxiv.org/abs/2106.06639 sha256:73e2e7f30df0…
`r419-metadata-cr`	Federated Learning with Buffered Asynchronous Aggregation (Crossref metadata) John Nguyen, Kshitiz Malik, Hongyuan Zhan, Ashkan Yousefpour, Michael Rabbat, Mani Malek, Dzmitry Huba	primary paper abstract_only	ID: https://proceedings.mlr.press/v151/nguyen22b.html sha256:c040b0bb8b85…

Ontology & Grounding Topology

Open in interactive network graph →

Curation Provenance

Curator: Antigravity

Method: ai_assisted

Model: Gemini

Created: 2026-10-19T08:22:00Z

Updated: 2026-10-19T08:22:00Z

Source File: data/curated/FedBuffBufferedAsynchronousFederatedAggregation.yaml · Schema: curated_concept.yaml

Explore across X-Mech

Related computational & biological Mechs: Embedding Map Network Graph ProteinTraitsMech (ESM-2 pLM) DUFMech (Domain pLM) TraitMech CultureMech All 14 X-Mech Repositories