FedACT: Resource-Aware FL Scheduling

• FedACT is a resource heterogeneity-aware scheduling approach for concurrent federated learning jobs on diverse edge devices, combining resource–demand alignment with participation fairness.
• It minimizes average job completion time while improving global model accuracy, achieving speedups up to 8.7× compared to baseline scheduling methods.
• By employing online per-round score computation with fairness regularization, FedACT effectively mitigates the challenges posed by non-IID data and system heterogeneity.

FedACT most specifically denotes a resource heterogeneity-aware device scheduling approach for concurrent federated learning in which multiple FL jobs share a common pool of heterogeneous edge devices. It is designed to minimize average job completion time (JCT) while preserving or improving global-model accuracy, especially under non-IID data, by combining resource–demand alignment with participation fairness in an online per-round scheduler (Islam et al., 11 Mar 2026). In the formulation studied, jobs run concurrently with equal priority, devices expose compute, memory, and network bandwidth availability, and each device may serve at most one job per round. The name is not fully unique across the federated learning literature, but in this sense it refers to the concurrent scheduling framework introduced in 2026.

1. Problem formulation and system model

FedACT is posed for a system with a set of $K$ edge devices $\mathcal{K}$ and a set of $M$ concurrent FL jobs $\mathcal{M}$, where each job denotes the training of one model (Islam et al., 11 Mar 2026). Devices are heterogeneous in compute, memory, and network bandwidth, periodically report current resource availability to the FL server, and hold local datasets for each job. Jobs submit minimum resource requirements per resource type and training hyperparameters, proceed in FL rounds, and use standard server aggregation.

For job $m$, device $k$ holds data $\mathcal{D}_k^m$ of size $D_k^m$, and the global dataset is $\mathcal{D}^m = \bigcup_k \mathcal{D}_k^m$ with size $D^m = \sum_k D_k^m$. The multi-job learning problem is written as

$\mathcal{K}$0

where

$\mathcal{K}$1

and $\mathcal{K}$2 (Islam et al., 11 Mar 2026).

Within each job, training follows a standard synchronous round-based loop. Each job $\mathcal{K}$3 has rounds $\mathcal{K}$4 and local epochs $\mathcal{K}$5; in the experiments, $\mathcal{K}$6. At round $\mathcal{K}$7, the server selects a fraction $\mathcal{K}$8 of devices, broadcasts the global model, devices train locally for $\mathcal{K}$9 epochs, and the server aggregates according to

$M$0

In the reported experiments, each job selects $M$1 of devices per round, i.e., $M$2 (Islam et al., 11 Mar 2026).

The motivating deployment scenario is explicitly system-level rather than single-task. Real deployments may host multiple models such as next-word prediction, image recommendation, and speech recognition at once across the same devices. Naive sequential execution or single-job schedulers extended to multi-job settings can underutilize devices, ignore device–job fit, and exacerbate data imbalance across devices, leading to prolonged overall training time and degraded accuracy (Islam et al., 11 Mar 2026). This suggests that FedACT is best understood not as a variant of FedAvg, but as a scheduling layer for shared, heterogeneous FL infrastructures.

2. Job completion time and heterogeneity model

The principal systems metric is the average job completion time. The paper defines JCT as the “wall-clock training time required for each job to reach its specified target accuracy or complete a predefined number of communication rounds,” and reports the mean over jobs as

$M$3

where $M$4 is the submission time and $M$5 is the time when the job first meets its target accuracy or loss $M$6, or finishes its maximum rounds $M$7 (Islam et al., 11 Mar 2026). Scheduling changes AvgJCT by determining which devices contribute to each job in each round, thereby altering the wall-clock progression of every job toward its stopping criterion.

FedACT models device–job compatibility along three resource dimensions: computational power, memory, and network bandwidth. Device resource values are normalized relative to each device’s total capacity before scoring. Each job declares minimum resource requirements along the same dimensions, and only jobs whose minimum requirements are satisfied by a device’s currently available resources are eligible for assignment on that device in the current round; oversubscription is excluded (Islam et al., 11 Mar 2026).

The execution-time model captures compute and communication heterogeneity through a shifted exponential distribution:

$M$8

where $M$9 is the execution time, $\mathcal{M}$0 is the number of local epochs for job $\mathcal{M}$1, $\mathcal{M}$2 is device $\mathcal{M}$3’s local data size for job $\mathcal{M}$4, and device parameters $\mathcal{M}$5 and $\mathcal{M}$6 jointly capture maximum capacity and fluctuation in compute and communication (Islam et al., 11 Mar 2026).

In this formulation, the scheduler is simultaneously responding to two forms of heterogeneity: instantaneous resource availability and statistical heterogeneity induced by non-IID local data. The former is reflected in JCT and straggler behavior, while the latter enters through participation fairness and its effect on model accuracy.

3. Alignment scoring and online scheduling

FedACT computes, for each eligible device–job pair at round $\mathcal{M}$7, a normalized score

$\mathcal{M}$8

where $\mathcal{M}$9 and $m$0 are empirically chosen weights, $m$1 is the resource alignment score, and $m$2 is the participation fairness score (Islam et al., 11 Mar 2026). The scoring function is the central mechanism by which the method trades off system efficiency and statistical balance.

The resource alignment term is computed as a weighted dot product between a device’s normalized available resource vector and a job’s normalized minimum demand vector; in the experiments, compute, memory, and bandwidth are equally weighted. Before this computation, the scheduler filters out all jobs whose minimum requirements cannot be fully met by the device’s currently available resources. The operational intuition is explicit: jobs that better utilize a device’s abundant resource receive higher alignment scores, and devices with greater availability in a resource favor jobs that demand more of that resource (Islam et al., 11 Mar 2026).

The fairness term regularizes repeated participation. Let $m$3 be the cumulative count of times device $m$4 has been scheduled for job $m$5 up to round $m$6. FedACT defines

$m$7

This penalizes deviations from the mean participation count for a given job: underrepresented devices receive higher fairness scores and overrepresented devices receive lower scores (Islam et al., 11 Mar 2026). The term is not a hard constraint; it is a regularizer inside the score. In non-IID regimes, its role is to counteract systematic over-selection of resource-strong devices and to diversify the training population.

The per-round decision can be expressed with binary variables $m$8 and the selected set $m$9. For each job $k$0, FedACT performs greedy maximization over not-yet-occupied devices:

$k$1

subject to eligibility and per-round device exclusivity (Islam et al., 11 Mar 2026). The paper describes the resulting scheduling plan as “optimal” in the sense of prioritizing higher-score devices under the stated constraints, but it does not provide a formal optimality proof.

Operationally, the scheduler repeats six steps: resource probing, eligibility filtering, score computation, top-$k$2 selection among not-yet-occupied devices, FL round execution, and fairness-state update. Jobs are executed “asynchronously in parallel,” while each individual job retains the standard synchronous FedAvg-style round structure (Islam et al., 11 Mar 2026). A plausible implication is that FedACT is most naturally deployed as an online control layer above multiple otherwise standard FL training loops.

4. Architecture, experimental design, and empirical behavior

The system architecture consists of a central FL server, multiple job managers, and a shared device pool. Jobs are submitted with minimum resource requirements; the server initializes each job’s global model, possibly random or pretrained, and runs per-job schedulers online. Devices periodically report compute, memory, and bandwidth availability; the server collects this metadata along with model updates, while sensitive data remain on-device (Islam et al., 11 Mar 2026).

The empirical study uses $k$3 devices, SGD for local training, and five datasets and five models: CIFAR-10, MNIST, EMNIST-Letters, EMNIST-Digits, Fashion-MNIST, together with LeNet-5, CNN variants, VGG-16, ResNet-18, and AlexNet. Two concurrent job groups are evaluated:

Group	Jobs
A	VGG-16 on CIFAR-10; CNN on EMNIST-Letters; LeNet-5 on EMNIST-Digits
B	ResNet-18 on CIFAR-10; CNN on Fashion-MNIST; AlexNet on MNIST

Model complexities are ordered as LeNet $k$4 CNN(A) $k$5 VGG/AlexNet $k$6 CNN(B) $k$7 ResNet (Islam et al., 11 Mar 2026). Data are tested under IID and non-IID conditions; the non-IID setting is formed by class-wise partitioning into 20 shards and assigning two class shards per device. The server hardware is reported as 4$k$8NVIDIA RTX A4000 GPUs, an Intel i9-10900X CPU, and 64 GB RAM (Islam et al., 11 Mar 2026).

The baselines span naive, heuristic, evolutionary, and model-based schedulers: Random, Greedy, Genetic, and MJ-FL (AAAI’22/TPDS’22) (Islam et al., 11 Mar 2026). Against these baselines, FedACT reduces average JCT by up to $k$9 and improves model accuracy by up to $\mathcal{D}_k^m$0 overall. In the IID setting, it converges substantially faster than baselines and often to higher final accuracy; for VGG on CIFAR-10 (IID), it attains up to $\mathcal{D}_k^m$1 higher final accuracy than Greedy and reaches the $\mathcal{D}_k^m$2 target up to $\mathcal{D}_k^m$3 faster than Random and Genetic, and $\mathcal{D}_k^m$4 faster than MJ-FL. Across jobs, it reduces overall time up to $\mathcal{D}_k^m$5 (Islam et al., 11 Mar 2026).

The non-IID setting is where the fairness term becomes most consequential. FedACT reports final accuracy improvements of $\mathcal{D}_k^m$6, $\mathcal{D}_k^m$7, $\mathcal{D}_k^m$8, and $\mathcal{D}_k^m$9 over Greedy, Genetic, Random, and MJ-FL, respectively, and reaches target accuracy with up to $D_k^m$0 and $D_k^m$1 reductions in average JCT for Group A and Group B (Islam et al., 11 Mar 2026). Gains are especially pronounced on complex jobs such as VGG and ResNet. The paper also compares against sequential single-job FedAvg execution and reports up to $D_k^m$2 faster wall-clock time to reach similar accuracy, including Group B non-IID speedups of $D_k^m$3 for ResNet at $D_k^m$4, $D_k^m$5 for CNN at $D_k^m$6, and $D_k^m$7 for AlexNet at $D_k^m$8 (Islam et al., 11 Mar 2026).

These results suggest that the benefit of FedACT does not arise solely from faster resource matching. The method’s empirical profile indicates that balanced device participation can materially improve convergence and final accuracy under statistical heterogeneity, rather than merely preserving them.

5. Robustness, hyperparameters, and limitations

FedACT’s two main hyperparameters are the alignment weight $D_k^m$9 and fairness weight $\mathcal{D}^m = \bigcup_k \mathcal{D}_k^m$0, both selected empirically based on prior runs and fine-tuned over short epochs (Islam et al., 11 Mar 2026). Increasing $\mathcal{D}^m = \bigcup_k \mathcal{D}_k^m$1 emphasizes resource alignment, accelerates convergence, and reduces JCT; increasing $\mathcal{D}^m = \bigcup_k \mathcal{D}_k^m$2 prioritizes fairness and improves accuracy under non-IID data. The method is reported to be stable under the chosen $\mathcal{D}^m = \bigcup_k \mathcal{D}_k^m$3 settings.

Several practical properties are emphasized. The shifted-exponential time model captures variability in compute and communication, and resource-aligned matching reduces straggler impact by assigning heavy jobs to capable devices. Per-round resource reporting makes the scheduler adaptive to fluctuating resource availability. Privacy is treated in the standard FL sense: only model updates and resource availability are shared, while sensitive data stay on-device. The framework is stated to be compatible with homomorphic encryption, secure aggregation, and differential privacy. The extra overhead is primarily score computation and sorting to form $\mathcal{D}^m = \bigcup_k \mathcal{D}_k^m$4, and this overhead is reported as negligible relative to the training benefits (Islam et al., 11 Mar 2026).

The paper does not provide formal proofs of optimality, JCT bounds, or convergence guarantees (Islam et al., 11 Mar 2026). Its use of “optimal” is therefore procedural rather than theorem-backed. The analysis also rests on several assumptions: accurate device resource reporting, negligible server aggregation time, equal job priority, and one job per device per round. The resource model covers compute, memory, and bandwidth, but not energy constraints. Future work identified in the paper includes asynchronous model updates and allowing devices to train multiple jobs concurrently without extra server-side aggregation delays (Islam et al., 11 Mar 2026).

A common misconception is that FedACT proves a globally optimal multi-job schedule. The paper does not claim such a proof. A more precise reading is that it proposes an online greedy scheduler whose empirical behavior is favorable under heterogeneous resources and non-IID data.

6. Naming ambiguity and other uses of “FedACT”

The label “FedACT” is not unique across the literature. In medical imaging, the 2023 paper “Federated Alternate Training (FAT): Leveraging Unannotated Data Silos in Federated Segmentation for Medical Imaging” states that FedACT can be regarded as the same approach as FAT, or a synonymous label emphasizing the Alternate Training mechanism in a federated regime (Mushtaq et al., 2023). In that usage, the method alternates between annotated silos, which optimize a supervised Dice-plus-cross-entropy segmentation objective, and unannotated silos, which perform teacher–student pseudo-labeling and EMA-based self-supervised learning with nnU-Net. That FedACT/FAT is therefore a federated semi-supervised segmentation framework rather than a concurrent job scheduler.

Several other arXiv works document adjacent naming confusions. The production FL framework FACT is explicitly noted as not being “FedACT”; the paper states that “FedACT” appears to be a misnomer or typo for FACT in that context (Weber et al., 2022). The optimization method FedAc, “Federated Accelerated Stochastic Gradient Descent,” likewise notes that “FedACT” does not appear in the paper and would almost surely be a typo or informal reference to FedAc (Yuan et al., 2020). The group-fairness framework FedFACT makes a similar clarification, treating “FedACT” as a naming variant or typo for FedFACT (Zhang et al., 4 Jun 2025). In reinforcement learning, “FedACT” is used at a higher level to denote the general idea of federated actor–critic, with SFAC presented as a concrete single-loop instantiation across heterogeneous environments (Zhu et al., 2024).

Accordingly, the most precise contemporary use of “FedACT” depends on context. In concurrent federated learning systems, it refers to the 2026 scheduler for multiple FL jobs over heterogeneous devices (Islam et al., 11 Mar 2026). In federated medical image segmentation, it may denote the alternate-training method otherwise named FAT (Mushtaq et al., 2023). The broader literature therefore requires explicit disambiguation whenever the term appears.