SplitLearning-B5G

Cleyber B. dos Reis¹, Maria do Rosário C. Ribeiro ¹, Antonio Oliveira-JR¹

¹Universidade Federal de Goiás - Goiânia, GO - Brazil

cleyber.bezerra@discente.ufg.br, rosario.ribeiro@ufg.br, antonio@inf.ufg.br

The repository hosts the full development carried out for the master’s dissertation Split Learning as an enabler of wireless networks for future generations, undertaken within the Academic Master’s Programme in Computer Science at the Institute of Informatics (INF), Federal University of Goiás (UFG), Goiânia.

Beyond serving as the dissertation’s experimental foundation, the repository also contains the implementation of SplitLearning-ns3, the network-aware Split Learning architecture introduced in the article SLArch: A Network Metric-aware Split Learning Architecture for B5G/6G Mobile Networks. This framework bridges wireless communication metrics — latency, throughput, jitter, packet loss rate, and energy consumption — with distributed training processes. In doing so, it enables reproducible and extensible experimentation in B5G/6G scenarios, providing both a research tool and a reference point for future studies.

Table of Contents

• Article Summary
  • Abstract
  • Baselines
  • Results
• Replicating the Experiment
  • Requirements
  • Preparing Environment
  • Run Experiments

Abstract

This paper introduces SplitLearning-ns3, a network-aware Split Learning (SL) architecture developed on top of ns3-ai for the NS-3 simulator (5G-LENA). More than a simple integration, SplitLearning-ns3 provides a reproducible and extensible framework that couples wireless communication metrics with distributed training processes. The experimental protocol subjects Convolutional Neural Network (CNN) training to realistic network dynamics—including latency, throughput, jitter, packet loss rate (PLR), and energy consumption—that characterise Beyond 5G (B5G) and 6G environments.

The results show that PLR is the dominant factor hindering convergence, whereas latency and throughput have a moderate influence, and jitter and energy consumption remain secondary though still measurable. Slice-level analysis further reveals that URLLC consistently ensures lower latency, while eMBB captures the largest share of throughput. These findings highlight the decisive importance of link reliability and energy awareness in evaluating the feasibility of SL for next-generation mobile networks, and motivate further research into resilient and energy-efficient learning at the wireless edge.

Back to TOC

Baselines

In this study, the baselines are defined by a reproducible experimental protocol implemented in SplitLearning-ns3. A single CNN model (MNIST) is trained across client–server partitions, with network dynamics injected from ns-3/5G-LENA. The baseline simulation considers two gNBs and 102 UEs operating under heterogeneous slices (URLLC, eMBB, and mMTC), with numerologies µ = 4 and µ = 2, a 100 MHz bandwidth, On–Off traffic, and transmission powers of 26 dBm (gNBs) and 13 dBm (UEs).

Both network metrics (latency, jitter, throughput, PLR, and energy consumption) and ML outcomes (validation accuracy and training time) are assessed jointly. Within this baseline, packet loss rate (PLR) emerges as the dominant constraint, outweighing delay and throughput, which contribute meaningfully only once link reliability is ensured.

Back to TOC

Results

Communication network and Split Learning environment

The proposed framework integrates network simulation with distributed training, enabling a joint analysis of key metrics such as latency, throughput, jitter, packet loss rate (PLR), energy consumption, and validation accuracy. Results are presented for heterogeneous slices (URLLC, eMBB, and mMTC).

Fig. 1. Latency distribution across slices (URLLC, eMBB, mMTC; 95% CI).

Fig. 2. Average throughput and stability across slices (95% CI).

Figures 1 and 2 show that URLLC consistently achieved the lowest latency due to prioritised scheduling, while eMBB dominated throughput allocation (≈79% of total capacity). Although mMTC contributed little in terms of throughput, it introduced traffic bursts that increased delay variability.

Fig. 3. Jitter dynamics under bursty mMTC traffic (95% CI).

Fig. 4. Validation accuracy versus Packet Loss Rate (PLR) (mean ± 95% CI).

Figures 3 and 4 highlight that jitter alone had a limited impact, but became critical when combined with PLR. Even small losses (PLR < 2%) caused a sharp reduction in validation accuracy, confirming reliability as a decisive factor for convergence.

Fig. 5. Average energy consumption across slices (95% CI).

Fig. 6. Normalised slice profiles across network metrics (↑ better for Throughput; ↓ preferable for Latency, Jitter, PLR, and Energy).

Figures 5 and 6 demonstrate that energy consumption remained stable across slices, with only minor variations, while validation accuracy was highly sensitive to network reliability. PLR degraded accuracy by up to −70.1%, overshadowing the effects of delay (+57.6%) and throughput (+12.2%).

Key insight: These results confirm that link reliability is the cornerstone for ensuring stable Split Learning in B5G/6G environments.

Back to TOC

Replicating The Experiment

Requirements

• GNU (>=8.0.0)

  command to know the version of KERNEL, GCC and GNU Binutils in the terminal

    cat /proc/version

• GCC (>=11.4.0)

  commands to know the GCC version in the terminal.

    gcc --version
    ls -l /usr/bin/gcc*

• CMAKE (>=3.24)

  command to know the version of CMAKE in the terminal.

    cmake --version

• python (>=3.11.5)

  commands to know the version of PYTHON in the terminal.

    python --version
    python3 --version

• ns-allinone (3.45) or ns-3-dev

Back to TOC

Preparing Environment

INSTALL GCC AND MAKE

    sudo apt update
    sudo apt install build-essential
    sudo apt install gcc-10 g++-10

    sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-10 100
    sudo update-alternatives --install /usr/bin/g++ g++ /usr/bin/g++-10 100

    gcc --version
    make --version

INSTALl THE CMAKE

download the file in the desired version: https://cmake.org/download/

extract it, access the folder and run the command:

    ./bootstrap && make && sudo make install

INSTALL THE GIT

    sudo apt-get install git -y

INSTALL THE NS-3.45

    git clone https://gitlab.com/nsnam/ns-3.45.git
    git checkout -b ns-3.45-release ns-3.45

CONFIGURE AND COMPILE NS-3

In the ns-3.45 folder use the commands:

    ./ns3 configure --enable-examples --enable-tests
    ./ns3 build

INSTALL THE NR-LENA

    sudo apt-get install libc6-dev
    sudo apt-get install sqlite sqlite3 libsqlite3-dev
    sudo apt-get install libeigen3-dev

    cd contrib
    git clone https://gitlab.com/cttc-lena/nr.git
    cd nr
    git checkout -b 5g-lena-v3.1.y origin/5g-lena-v3.1.y

INSTALL THE NS3-IA

    sudo apt install python3-pip
    
    pip install tensorflow==2.17.0
    pip install cloudpickle==1.2.0
    pip install pyzmq
    pip install protobuf==3.20.3
    pip install tensorflow-estimator==2.15.0
    pip install tensorboard==2.17.0

    pip install numpy==1.18.1
    pip install Keras==3.2.0
    pip install Keras-Applications==1.0.8
    pip install Keras-Preprocessing==1.1.2

    pip install matplotlib==3.3.2
    pip install psutil==5.7.2

    cd contrib/
    git clone https://github.com/hust-diangroup/ns3-ai.git

    cd ns3-ai/py_interface
    pip3 install . --user

ENABLE MODULES

    ./ns3 configure --enable-modules=nr,internet-apps,flow-monitor,config-store,buildings,applications,network,core,wifi,energy,spectrum,propagation,mobility,antenna

DOWNLOAD PROJECT

inside the ns-3.45/scratch folder use the commands:

git clone https://github.com/LABORA-INF-UFG/SplitLearning-B5G.git

Back to TOC

Run Experiments

EXECUTE PROJECT

inside the Split Learning folder run the script file.

    ./simulator_ns3.sh

Back to TOC