基于AirComp的分布式CNN推理资源调度研究

刘乔寿; 邓义锋; 胡昊南; 杨振巍

doi:10.11999/JEIT241022

基于AirComp的分布式CNN推理资源调度研究

doi: 10.11999/JEIT241022 cstr: 32379.14.JEIT241022

刘乔寿^{1, 2, 3},
邓义锋^{1, 2, 3, ,},
胡昊南¹,
杨振巍^{1, 2, 3}

1.
重庆邮电大学通信与信息工程学院重庆 400065
2.
先进网络与智能互联技术重庆市高校重点实验室重庆 400065
3.
泛在感知与互联重庆重点实验室重庆 400065

基金项目: 重庆市教委科学技术研究计划重大项目(KJZD-M202400602, KJZD-K202200604)，重庆市自然科学基金面上项目(CSTB2024NSCQ-MSX0731)

详细信息

作者简介:
刘乔寿：男，副教授，博士，研究方向为5G超密集网络干扰协调、云边协同智能计算、FPGA智能算法加速、物联网系统及终端设备开发

邓义锋：男，硕士生，研究方向为空中计算和分布式推理

胡昊南：男，副教授，博士，研究方向为5G/B5G通信、5G NR-U蜂窝和WiFi共存网络、边缘计算

杨振巍：男，硕士生，研究方向为计算机视觉

通讯作者:
邓义锋　929803460@qq.com

中图分类号: TN929
计量
- 文章访问数: 129
- HTML全文浏览量: 108
- PDF下载量: 16
- 被引次数: 0
出版历程
- 收稿日期: 2024-11-14
- 修回日期: 2025-03-16
- 网络出版日期: 2025-03-25
- 刊出日期: 2025-07-22

Research on Resource Scheduling of Distributed CNN Inference System Based on AirComp

LIU Qiaoshou^{1, 2, 3},
DENG Yifeng^{1, 2, 3
, ,},
HU Haonan¹,
YANG Zhenwei^{1, 2, 3}

1.
School of Communications and Information Engineering, Chongqing University of Posts and Telecommunications Chongqing 400065, China
2.
Advanced Network and Intelligent Connection Technology Key Laboratory of Chongqing Education Commission of China Chongqing 400065, China
3.
Chongqing Key Laboratory of Ubiquitous Sensing and Networking Chongqing 400065, China

Funds: The Major Project of Science and Technology Research Program of Chongqing Municipal Education Commission (KJZD-M202400602, KJZD-K202200604), The General Project of Chongqing Natural Science Foundation (CSTB2024NSCQ-MSX0731)

摘要

摘要: 在传统AirComp系统中，汇聚节点接收到来自不同发送端的信号相位是否严格对齐将直接影响AirComp的计算精度，将AirComp引入分布式联邦学习和分布式推理系统中，由于相位对齐问题造成的计算误差则会导致模型训练精度和推理精度下降。目前，现有的AirComp分布式联邦学习和分布式推理系统，无论在训练还是推理过程中，基本上都未考虑信道对模型性能的影响，导致其推理精度远低于本地训练和推理的结果，这一点在低信噪比时表现得尤为突出。该文提出了一种MOSI-AirComp系统，其中同一轮参与计算的发射信号来自同一节点，因此可以忽略信号的相位对齐问题。此外，该文设计了一种双支路训练模型，上支路基于原始模型的基础上添加Loss层模拟信道干扰，而下支路保持原始的网络模型结构用于推理任务，以实现更好的抗衰落和抗噪声能力。该文还提出了一种基于权重的功率控制方案和路径选择算法，根据节点间距离和模型权重选择最优的传输回路，并将模型权重作为功率控制因子的一部分来调节传输功率，以此实现卷积过程中的乘法操作，同时利用AirComp的叠加特性完成加法操作，从而实现空中卷积。仿真结果证明了MOSI-AirComp系统的有效性。与传统模型相比，双支路训练模型在小尺度衰落场景下，MNIST数据集和CIFAR10数据集在不同信噪比下的推理精度分别提高了2%～18%和0.4%～11.2%。
- 空中计算(AirComp) /
- 分布式推理 /
- 卷积神经网络(CNN) /
- 功率控制
Abstract: Objective In traditional AirComp systems, the computational accuracy is directly affected by the alignment of received signal phases from different transmitters. When applied to distributed federated learning and distributed inference systems, phase misalignment can introduce computational errors, reducing model training and inference accuracy. This study proposes the MOSI-AirComp system, in which transmitted signals in each computation round originate from the same node, thereby eliminating signal phase alignment issues. Methods (1) A dual-branch training model is proposed, increasing network complexity only during training. The traditional model is extended to a dual-branch structure, where the lower branch retains the original model, and the upper branch incorporates additional loss layers for training. (2) An MOSI-AirComp-based weight-power control scheme is introduced. Each node is equipped with multiple transmitting antennas and a single receiving antenna. Pre-trained model weights are offloaded to task nodes as part of the power control factor, which adjusts transmission power during inference. This optimization enhances signal amplitude for convolution operations while reducing computation time. Since data transmission originates from the same node, phase alignment issues are avoided. AirComp integrates signals from multiple antennas for convolution summation, enabling airborne convolution. (3) A TSP-based node selection algorithm is proposed, using weight mean and path as evaluation parameters to determine the optimal transmission path, ensuring efficient data transmission. Results and Discussions Compared to the traditional network model, the dual-branch training model significantly improves inference accuracy under small-scale fading. For the MNIST and CIFAR-10 datasets, accuracy increases by 2%～18% and 0.4%～11.2% under different SNR values (Fig. 5 and Fig. 6). The MSE decreases by 0.056～0.154 and 0.047～0.23 under different maximum node power budgets (Fig. 7). In noise-only scenarios, inference accuracy improves by 0.7%～5.5% and 0.3%～7.1% under different SNR values (Fig. 5 and Fig. 6), while MSE decreases by 0.035～0.152 and 0.056～0.253 under different maximum node power budgets (Fig. 8). Conclusions An MOSI-AirComp system is proposed to address the phase alignment issue inherent in traditional AirComp scenarios. The system enables airborne convolution through a power control scheme and enhances the traditional network model with a dual-branch structure. The upper branch simulates multiplicative Rayleigh fading using loss layers and incorporates model data into the convolution layer output of the lower branch to simulate additive noise effects. To account for node limitations in IoT networks, a model-weight-improved Traveling Salesman Problem (TSP) node selection algorithm is proposed. Future advancements in AirComp deployment for distributed computing and communication frameworks hold promise, particularly with the rapid development of 6G and IoT.
- AirComp /
- Distributed inference /
- Convolutional Neural Network (CNN) /
- Power control

HTML全文

图 1 双支路训练模型

下载: 全尺寸图片幻灯片

图 2 传统Aircmop系统

下载: 全尺寸图片幻灯片

图 3 MOSI-AirComp系统

下载: 全尺寸图片幻灯片

图 4 节点选择算法流程图

下载: 全尺寸图片幻灯片

图 5 Lenet5网络下不同SNR的测试精度

下载: 全尺寸图片幻灯片

图 6 VGGNet16网络下不同SNR的测试精度

下载: 全尺寸图片幻灯片

图 7 Lenet5网络下不同$ {P_{\max }} $的MSE

下载: 全尺寸图片幻灯片

图 8 VGGNet16网络下不同$ {P_{\max }} $的MSE

下载: 全尺寸图片幻灯片

图 9 不同网络在不同的最大环路数下的推理延迟

下载: 全尺寸图片幻灯片

1 基于权重的节点选择算法

输入：节点总数$ K $、模型卷积层层数$ M $
输出：环路集合${c_m}$或最优的环路集合${T_{{\text{opt}}}}$
(1) 初始化节点位置信息${d_{i,j}}$、各层卷积的权重均值$ {\bar w_i} $、环路集　合${c_m}$、初始节点$ j $
(2) if $K \ge M$ then
(3) 　for $ m = 1,2, \cdots ,M $do
(4) 　计算第$ j $个节点到所有未访问节点距离
(5) 　节点$ j $搜索所有未访问的最短下一节点路径信息${d_{j,{\text{next}}}}$
(6) 　根据式(20)，记录 ${c_m} \leftarrow {c_m} + {\varepsilon _j}$
(7) 　更新$j \leftarrow $ next
(8) 　end for
(9) return ${c_m}$
(10) end if
(11) else if $K < M$ then
(12) for $i = 1,2, \cdots, K$ do
(13) 　if $ i > 0 $ do $i \leftarrow i - 1$
(14) 　　计算第$ j $个节点到所有未访问节点距离
(15) 　　节点$ j $搜索未访问的最短下一节点路径信息${d_{j,{\text{next}}}}$
(16) 　　根据式(20)，记录$ {c_i} \leftarrow {c_i} + {\varepsilon _j} $
(17) 　　更新$j \leftarrow {\text{next}}$
(18) 　end if
(19) end for
(20) 记录${c_i}$到环路集合$C$中，$C \leftarrow C + {c_i}$
(21) 根据式(22)、$a = \left[ {m/K} \right]$，计算出最优的环路集合${T_{{\text{opt}}}}$
(22) return ${T_{{\text{opt}}}}$
(23) end if

下载: 导出CSV

表 1 通信设置

设置项	值
网络最大范围(m×m)	100×100
最大节点功率预算(mW)	1
路径损耗指数	3
发射/接收天线数	9/1
SNR(dB)	25
节点总数	15

下载: 导出CSV

参考文献(16)

[1]	XIAO Jinjun, CUI Shuguang, LUO Zhiquan, et al. Linear coherent decentralized estimation[J]. IEEE Transactions on Signal Processing, 2008, 56(2): 757–770. doi: 10.1109/TSP.2007.906762.
[2]	GASTPAR M. Uncoded transmission is exactly optimal for a simple Gaussian "sensor" network[J]. IEEE Transactions on Information Theory, 2008, 54(11): 5247–5251. doi: 10.1109/TIT.2008.929967.
[3]	BUCK R C. Approximate complexity and functional representation[J]. Journal of Mathematical Analysis and Applications, 1979, 70(1): 280–298. doi: 10.1016/0022-247X(79)90091-X.
[4]	GOLDENBAUM M, BOCHE H, and STAŃCZAK S. Nomographic functions: Efficient computation in clustered Gaussian sensor networks[J]. IEEE Transactions on Wireless Communications, 2015, 14(4): 2093–2105. doi: 10.1109/TWC.2014.2380317.
[5]	ABARI O, RAHUL H, and KATABI D. Over-the-air function computation in sensor networks[EB/OL]. arXiv: 1612.02307, 2016. https://arxiv.org/abs/1612.02307.
[6]	WANG Zhibin, ZHAO Yapeng, ZHOU Yong, et al. Over-the-air computation for 6G: Foundations, technologies, and applications[J]. IEEE Internet of Things Journal, 2024, 11(14): 24634–24658. doi: 10.1109/JIOT.2024.3405448.
[7]	LEE Chuangzheng, BARNES L P, and ÖZGÜR A. Over-the-air statistical estimation[J]. IEEE Journal on Selected Areas in Communications, 2022, 40(2): 548–561. doi: 10.1109/JSAC.2021.3118412.
[8]	WANG Zhibin, ZHAO Yapeng, ZHOU Yong, et al. Over-the-air computation for 6G: Foundations, technologies, and applications[EB/OL]. arXiv: 2210.10524, 2022. https://arxiv.org/abs/2210.10524.
[9]	SIGG S, JAKIMOVSKI P, and BEIGL M. Calculation of functions on the RF-channel for IoT[C]. 2012 3rd IEEE International Conference on the Internet of Things, Wuxi, China, 2012: 107–113. doi: 10.1109/IOT.2012.6402311.
[10]	CHEN Jiale, VAN LE D, TAN Rui, et al. Split convolutional neural networks for distributed inference on concurrent IoT sensors[C]. 2021 IEEE 27th International Conference on Parallel and Distributed Systems (ICPADS), Beijing, China, 2021: 66–73. doi: 10.1109/ICPADS53394.2021.00014.
[11]	MAO Jiachen, CHEN Xiang, NIXON K W, et al. MoDNN: Local distributed mobile computing system for Deep Neural Network[C]. Design, Automation & Test in Europe Conference & Exhibition (DATE), Lausanne, Switzerland, 2017: 1396–1401. doi: 10.23919/DATE.2017.7927211.
[12]	SANCHEZ S G, REUS-MUNS G, BOCANEGRA C, et al. AirNN: Over-the-air computation for neural networks via reconfigurable intelligent surfaces[J]. IEEE/ACM Transactions on Networking, 2023, 31(6): 2470–2482. doi: 10.1109/TNET.2022.3225883.
[13]	ZENG Liekang, CHEN Xu, ZHOU Zhi, et al. CoEdge: Cooperative DNN inference with adaptive workload partitioning over heterogeneous edge devices[J]. IEEE/ACM Transactions on Networking, 2021, 29(2): 595–608. doi: 10.1109/TNET.2020.3042320.
[14]	FAN Shaoshuai, NI Wei, TIAN Hui, et al. Carrier phase-based synchronization and high-accuracy positioning in 5G new radio cellular networks[J]. IEEE Transactions on Communications, 2022, 70(1): 564–577. doi: 10.1109/TCOMM.2021.3119072.
[15]	YOU Lizhao, ZHAO Xinbo, CAO Rui, et al. Broadband digital over-the-air computation for wireless federated edge learning[J]. IEEE Transactions on Mobile Computing, 2024, 23(5): 5212–5228. doi: 10.1109/TMC.2023.3304652.
[16]	DONG Ying, HU Haonan, LIU Qiaoshou, et al. Modeling and performance analysis of over-the-air computing in cellular IoT networks[J]. IEEE Wireless Communications Letters, 2024, 13(9): 2332–2336. doi: 10.1109/LWC.2024.3407749.