PaperlayerView exact revision on arXiv

Source-linked AI summary

AutoFL: Enabling Heterogeneity-Aware Energy Efficient Federated Learning

Young Geun Kim, Carole-Jean Wu

arXiv:2107.08147v1cs.LGcs.DC

TL;DR

At the edge, FL efficiency is challenged by data and system heterogeneity and stochastic runtime effects. AutoFL uses reinforcement learning to select participants and their execution targets; real-system experiments report higher energy efficiency for both individual devices and the participant cluster while satisfying accuracy requirements.

  • Problem

    FL must choose participants and per-device execution targets amid system and data heterogeneity and stochastic runtime effects; random participant selection leaves energy-efficiency and convergence co-optimization opportunity.

    Source: §1 · pp. 1–2

  • Method

    AutoFL uses reinforcement learning to select participants and their execution targets for each round, accounting for heterogeneity and runtime variance while maximizing energy efficiency and satisfying accuracy targets.

    Source: §1 · p. 2, §4.1 · pp. 6–7

  • Results

    4.7x and 5.2x higher energy efficiency for individual participant devices and the cluster of all participating devices, respectively, while satisfying the accuracy requirement.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Real-system experiments demonstrate AutoFL's feasibility across a variety of FL use cases while satisfying accuracy requirements.

    Source: §1 · p. 2

    Energy estimation covers CPUs and GPUs because of limited programmability of on-device training; similar practice is suggested for other processors.

    Source: §4.1 · p. 6

Abstract

from arXiv · show

Federated learning enables a cluster of decentralized mobile devices at the edge to collaboratively train a shared machine learning model, while keeping all the raw training samples on device. This decentralized training approach is demonstrated as a practical solution to mitigate the risk of privacy leakage. However, enabling efficient FL deployment at the edge is challenging because of non-IID training data distribution, wide system heterogeneity and stochastic-varying runtime effects in the field. This paper jointly optimizes time-to-convergence and energy efficiency of state-of-the-art FL use cases by taking into account the stochastic nature of edge execution. We propose AutoFL by tailor-designing a reinforcement learning algorithm that learns and determines which K participant devices and per-device execution targets for each FL model aggregation round in the presence of stochastic runtime variance, system and data heterogeneity. By considering the unique characteristics of FL edge deployment judiciously, AutoFL achieves 3.6 times faster model convergence time and 4.7 and 5.2 times higher energy efficiency for local clients and globally over the cluster of K participants, respectively.

1. INTRODUCTION

Federated learning keeps raw training data on devices, but choosing participants and execution targets efficiently remains difficult amid system and data heterogeneity and stochastic runtime variance. AutoFL uses reinforcement learning to make these choices and improves energy efficiency while maintaining training accuracy.

1. INTRODUCTION

  • Random participant selection leaves substantial room to co-optimize energy efficiency and model convergence for each aggregation round.The paper quantifies the opportunity as 5.4 times for energy efficiency and 4.2 times for model convergence.Source: §1 · p. 1, Figure 1 · p. 1
  • Device, data, and runtime heterogeneity complicate participant selection: stragglers gate round time, while non-IID data can defer convergence and degrade model quality.Prior approaches described in the paper address these challenges separately, and none considers both data and system heterogeneity with runtime variance.Source: §1 · p. 2
  • Energy-efficient FL could support on-device training anytime, with better model quality and user experience.The paper notes that prior work often assumes smartphones are plugged into wall power during FL.Source: §1 · p. 2
  • AutoFL uses reinforcement learning to select K participants and each participant’s execution target per aggregation round, adapting to system and data heterogeneity and runtime variance.It observes model characteristics, global parameters, and device profiles, then feeds measured decisions back into subsequent learning.Source: §1 · p. 2
  • 4.7x and 5.2x average energy-efficiency gains for individual smartphones and the full participant cluster, respectively, while maintaining training accuracy.The real-system evaluation compares AutoFL with random selection.Source: §1 · p. 2

2. BACKGROUND

Federated learning trains a shared model through repeated local training and server aggregation while keeping raw data on devices. In realistic deployments, device and data heterogeneity and stochastic runtime conditions make execution and convergence variable.

2.1 Federated Learning

  • Each round, a server selects K of N devices, sends them the global model, and averages their returned gradients after local training with batch size B for E epochs.The rounds repeat until the desired accuracy is reached.Source: §2.1 · p. 3

2.2 Consideration for Realistic Execution Environment

  • The FL optimization space spans large-scale decentralized training, device-performance differences, available co-processors, and heterogeneous local data.Different device capabilities and data distributions shape execution and model convergence.Source: §2.2 · p. 3, Figure 3 · p. 3
  • System heterogeneity creates performance gaps that produce stragglers, while interference, temperature, and network conditions make computation and communication times dynamic.The training time of each round can be gated by its slowest participant.Source: §2.2 · p. 3
  • Non-IID local data can significantly deteriorate FL convergence and model quality.Not all participating devices have IID training samples.Source: §2.2 · p. 3

3. MOTIVATION

The best participant cluster changes with FL parameters, model characteristics, runtime conditions, and data heterogeneity. These dependencies motivate selecting devices and execution targets in light of the current use case and execution environment.

3. MOTIVATION

  • The motivation examines energy efficiency, convergence time, and accuracy as the three axes of FL system characterization.Source: §3 · p. 3

3.1 Impact of FL Global Parameters and NN Characteristics

  • Across FL global parameter settings S1–S4, the optimal CNN-MNIST cluster changes from C1 to C2, C3, and C4, respectively.With more computation, high-end devices improve training time; with less computation, lower power consumption of mid- and low-end devices can improve energy efficiency.Source: §3.1 · pp. 3–4
  • For LSTM-Shakespeare, the optimal clusters across S1–S4 are C3, C4, C5, and C5, unlike CNN-MNIST’s C1, C2, C3, and C4.The smaller performance gap among device categories for LSTM-Shakespeare lets lower power consumption amortize performance loss.Source: §3.1 · p. 4

3.2 Impact of Runtime Variance

  • The optimal cluster shifts from C3 in ideal conditions to C1 with on-device interference and C5 with weak network signal; C3 achieves 3.2x higher energy efficiency than baseline C0.Poor signal increases communication time and energy by 4.3x on average, making low-power devices beneficial for energy efficiency.Source: §3.2 · p. 4, Figure 5 · p. 4

3.3 Impact of Data Heterogeneity

  • A >85% energy-efficiency gap separates ideal device selection from suboptimal selection with non-IID data, which also significantly increases convergence time.Figure 6 compares model quality and energy efficiency across levels of data heterogeneity.Source: §3.3 · p. 4, Figure 6 · p. 4

4. AUTOFL

AutoFL uses reinforcement learning to choose participating devices and their execution targets each aggregation round, adapting to FL settings, device conditions, data heterogeneity, and runtime variation. Its design models state, action, and reward for this optimization, then updates per-device Q-tables from measured outcomes.

4. AUTOFL

  • AutoFL selects participant devices and per-device execution targets each aggregation round to improve energy efficiency while satisfying accuracy requirements.Its decisions use FL configurations, device resource and network states, and per-device data-class counts.Source: §4 · pp. 4–5
  • AutoFL uses Q-learning with epsilon-greedy exploration, balancing exploratory actions against selecting actions with the highest reward.The authors motivate reinforcement learning for stochastic variance and cite faster training and inference, sample efficiency, and prediction accuracy as advantages over other statistical methods.Source: §4 · pp. 4–5

4.1 AutoFL Reinforcement Learning Design

  • The state captures neural-network layer types and global batch-size, local-epoch, and participant-count settings, alongside device resource use, network bandwidth, and data-class counts.DBSCAN discretizes continuous state features for the Q-learning lookup table.Source: §4.1 · p. 6
  • Actions jointly cover global participant selection and each selected device’s execution target, including CPU DVFS settings to use performance slack for energy savings.Available targets include CPUs, GPUs, and DSPs.Source: §4.1 · p. 6
  • The reward combines estimated local and cluster energy with test accuracy and accuracy improvement, and penalizes actions that fail to improve accuracy.The authors report a 7.3% mean absolute percentage error for energy estimates, which they describe as low enough to identify optimal participants and execution targets.Source: §4.1 · pp. 6–7

4.2 AutoFL Implementation Detail

  • AutoFL initializes per-device Q-tables, explores randomly with probability ε, otherwise selects high-value participants and targets, then updates tables from measured rewards.After learning converges, it uses the per-device tables to select participants and actions for observed states; ties among devices are broken randomly.Source: §4.2 · p. 7

5. EXPERIMENTAL METHODOLOGY

The evaluation uses 200 emulated mobile systems, representative FL workloads, and measured device profiles to assess AutoFL against baseline and prior designs. Its design evaluates participant selection and execution targets, including DVFS, under varied workloads, runtime conditions, and data distributions.

5.1 System Measurement Infrastructure

  • The evaluation uses an edge-cloud FL system with 200 mobile devices: 30 high-end, 70 mid-end, and 100 low-end, with power measured directly during on-device training on three smartphones.Source: §5.1 · pp. 7–8
  • The comparisons include FedAvg-Random, power- and performance-oriented selection, optimal participant selection, OFL, FedNova, and FEDL.Oparticipant accounts for heterogeneity and runtime variance in participant selection; OFL also considers available on-device coprocessors.Source: §5.1 · p. 8

5.2 Workloads and Execution Scenarios

  • The evaluation covers CNN-MNIST, LSTM-Shakespeare, and MobileNet-ImageNet workloads, with runtime variance and four levels of data heterogeneity represented in execution scenarios.Runtime variance is emulated through co-running applications and Gaussian network delay; data distributions range from Ideal IID to Non-IID (100%).Source: §5.2 · p. 8

5.3 AutoFL Design Specification

  • AutoFL uses two-level actions for participant-cluster and execution-target selection, augmenting targets with device DVFS settings to exploit straggler-related performance slack.The evaluation uses a learning rate of 0.9 and discount factor of 0.1, selected through hyperparameter comparisons.Source: §5.3 · pp. 8–9

6. EVALUATION RESULTS AND ANALYSIS

Across evaluated workloads, AutoFL improves energy efficiency and training outcomes by selecting participants and execution targets while adapting to runtime variance and data heterogeneity. Comparisons also show benefits against prior methods, with exploration and implementation overheads quantified.

6.1 Result Overview

  • 4.3x, 3.2x, and 2.0x higher average FL energy efficiency for CNN-MNIST, LSTM-Shakespeare, and MobileNet-ImageNet, respectively, versus FedAvg-Random, Power, and Performance.AutoFL also shows better training accuracy; Figure 8 compares performance-per-watt, convergence time, and training accuracy, with the first two normalized to FedAvg-Random.Source: §6.1 · p. 9, Figure 8 · p. 9
  • 3.5x, 2.9x, and 1.8x better training time per round than FedAvg-Random, Power, and Performance, respectively, accompany participant selection that reduces straggler slack.Execution-target selection and DVFS improve energy efficiency by an average of 19.8% over Oparticipant, with slightly higher convergence time.Source: §6.1 · p. 9

6.2 Adaptability and Accuracy Analysis

  • Across four global-parameter settings and, under runtime variance, different execution environments, AutoFL consistently improves training or convergence time and energy efficiency for federated learning.Figure 8 reports 4.0x, 3.7x, and 5.1x higher energy efficiency over FedAvg-Random for CNN-MNIST, LSTM-Shakespear, and MobileNet-ImageNet, respectively.Source: Figure 9 · p. 9, Figure 8 · p. 9, Figure 10 · p. 10
  • In the presence of runtime variance, AutoFL improves average energy efficiency by 5.1x, 6.9x, and 2.6x compared with FedAvg-Random, Power, and Performance.It improves convergence time by 3.4x, 3.3x, and 2.3x against those baselines, respectively, and achieves nearly similar energy efficiency, convergence time, and training accuracy to OFL.Source: §6.2 · p. 10, Figure 14 · p. 11
  • With data heterogeneity, AutoFL improves energy efficiency by 7.4x, 5.5x, and 4.3x versus FedAvg-Random, Power, and Performance, while improving convergence time and training accuracy.In Non-IID (75%) and Non-IID (100%) scenarios, CNN-MNIST does not converge within 1000 rounds under baseline settings; AutoFL achieves near-optimal energy efficiency, convergence time, and model quality.Source: §6.2 · p. 10, Figure 11 · p. 10

6.3 Comparison with Prior Work

  • AutoFL achieves 49.8% and 39.3% higher average energy efficiency than FedNova and FEDL, respectively, and raises performance-per-watt by 62.7% and 48.8% over them under stochastic variance.FedNova and FEDL allow partial updates but use random participant selection and do not exploit other execution targets.Source: §6.3 · p. 11, Figure 13 · p. 11, Figure 14 · p. 11

6.4 Overhead Analysis

  • After Q-table reward convergence, usually within 50–80 aggregation rounds, AutoFL achieves 5.2x average energy efficiency improvement for the entire FL relative to FedAvg-Random.Before convergence, it has 28.3% lower average energy efficiency than OFL but still saves 52.1% energy against FedAvg-Random; shared Q-tables reduce training overhead by 29.3% while prediction accuracy falls 2.7%.Source: §6.4 · p. 11, Figure 15 · p. 12
  • Training per-device Q-tables costs 531.5 µs on average, while AutoFL requires 80MB total memory for 200 devices and has 93.8% overall prediction accuracy.Sharing Q-tables across device categories speeds learning; the reported timing and energy-efficiency overheads from inference misprediction are 5.6% and 8.8%, respectively.Source: §6.4 · pp. 11–12

7. RELATED WORK

Prior FL optimization work addresses stragglers, data heterogeneity, or energy use, but the related-work discussion identifies a gap in handling these conditions together with stochastic runtime variance. AutoFL is presented as a reinforcement-learning approach to that combined setting.

Related work on stragglers

  • Prior straggler approaches exclude slow devices or use asynchronous gradient updates, while CPU-frequency adjustment for power saving can substantially increase overall training time.Excluding stragglers or asynchronous updates can also result in accuracy loss because of insufficient gradient updates.Source: §7 · p. 12

Related work on data heterogeneity

  • Prior approaches to data heterogeneity exclude non-IID updates, warm up models with shared data, share data across devices, or use FedProx partial updates, but do not jointly handle runtime variance and participant heterogeneity.The paper notes AutoFL can be used with FedProx to improve device selection.Source: §7 · p. 12

Energy optimization for FL

  • AutoFL is described as the first work demonstrating energy-efficient FL execution amid system and data heterogeneity with realistic performance uncertainty, using reinforcement learning to select participants and execution targets.The discussion contrasts this need with prior work that often assumes FL runs only when smartphones are plugged into wall power.Source: §7 · p. 12

8. CONCLUSION

The conclusion presents AutoFL as an adaptive, lightweight framework that learns participant and execution-target choices for energy-constrained mobile FL. In the evaluated emulated environment, it improves energy efficiency while also improving convergence time and accuracy.

8. CONCLUSION

  • AutoFL improves FL energy efficiency by 5.2x on average over random selection while improving convergence time and accuracy.The conclusion describes the evaluation as representative FL workloads deployed in an emulated mobile-cloud environment using off-the-shelf systems.Source: §8 · p. 12
Loading 2107.08147v1…