1. Introduction

• In this use case, we utilize the "Cell Metrics"(RRU.PrbUsedDl) dataset provided by the O-RAN SC SIM space, which includes synthetic data generated by a simulator, with all data recorded in Unix timestamp format.

  • Download data: https://nexus3.o-ran-sc.org/repository/datasets/CellReports.csv

• The model training process is carried out on the O-RAN SC AI/ML Framework, including GPU support, and considers both traditional machine learning (ML) and deep learning (DL) approaches. For ML models, we use Random Forest and Support Vector Regression (SVR), while for DL models, we employ RNN, LSTM, and GRU architectures.

• By managing the ON/OFF state of cells through traffic forecasting, we can reduce power consumption. Additionally, if the AI/ML models used for forecasting are operated in an eco-friendly manner, further power savings can be achieved. In this use case, we measure the carbon emissions and energy consumption during the cell traffic forecasting process using AI/ML to ensure that the forecasting model is not only effective but also environmentally sustainable.

...

Configuring GPU Usage in a Machine Learning Pipeline

• Step 1. Install the nvidia-container-toolkit

  Code Block
  language bash
  curl -fsSL https://nvidia.github.io/libnvidia-container/gpgkey | sudo gpg --dearmor -o /usr/share/keyrings/nvidia-container-toolkit-keyring.gpg \ && curl -s -L https://nvidia.github.io/libnvidia-container/stable/deb/nvidia-container-toolkit.list | \ sed 's#deb https://#deb [signed-by=/usr/share/keyrings/nvidia-container-toolkit-keyring.gpg] https://#g' | \ sudo tee /etc/apt/sources.list.d/nvidia-container-toolkit.list sudo apt-get update && sudo apt-get install -y nvidia-container-toolkit

• Step 2. Configure containerd

  Code Block
  language bash
  sudo nvidia-ctk runtime configure --runtime=containerd

  Code Block
  language bash
  sudo vim /etc/containerd/config.toml

  Code Block

  [plugins."io.containerd.grpc.v1.cri".containerd] default_runtime_name = "nvidia" # Change to "nvidia" # Additional configurations are omitted [plugins."io.containerd.grpc.v1.cri".containerd.runtimes.nvidia] [plugins."io.containerd.grpc.v1.cri".containerd.runtimes.nvidia.options] BinaryName = "/usr/bin/nvidia-container-runtime" # Additional configurations are omitted # Include the following content below [plugins."io.containerd.grpc.v1.cri".containerd.runtimes.nvidia.env] LD_LIBRARY_PATH = "/usr/lib/x86_64-linux-gnu:/usr/local/cuda/lib64"

  Code Block
  language bash
  # Restart containerd service sudo systemctl restart containerd

• Step 3. Install the nvidia-device-plugin

  Code Block
  language bash
  kubectl create -f https://raw.githubusercontent.com/NVIDIA/k8s-device-plugin/v0.16.2/deployments/static/nvidia-device-plugin.yml

• Step 4. Build the traininghost/pipelinegpuimage image

  • We built a new base image that recognizes the configured GPU to enable proper GPU usage in the ML pipeline components

  • To build the required image, you can refer to the provided Dockerfile and requirement.txt at the following link, or modify the pipeline image available in the existing aimlfw-dep

  Code Block
  language bash
  sudo buildctl --addr=nerdctl-container://buildkitd build \ --frontend dockerfile.v0 \ --opt filename=Dockerfile.pipeline_gpu \ --local dockerfile=pipeline_gpu \ --local context=pipeline_gpu \ --output type=oci,name=traininghost/pipelinegpuimage:latest | sudo nerdctl load --namespace k8s.io

• Step 5. Verify GPU usage with nerdctl

  • If the output is similar to the one below, the GPU setup is complete.

  Code Block
  language bash
  sudo nerdctl run -it --rm --gpus all --namespace k8s.io -p 8888:8888 -v $(pwd):/app_run traininghost/pipelinegpuimage:latest /bin/bash -c "nvidia-smi"

...

• Step 1. Download the pipeline script(pipeline.ipynb) provided in the “File List”

• Step 2. Modify the pipeline script to satisfy your own requirements

  • Set data features required for model training (using FeatureStoreSdk)

    • We used the RRU_PrbUsedDl column from the Viavi Dataset
      we extracted data at 30-minute intervals to train the model, focusing on capturing meaningful patterns in traffic data.

  • Write a TensorFlow-based AI/ML model script

    • We used LSTM(Long Short-Term Memory) model to predict downlink traffic

    • You can add other model prediction accuracy(e.g. RMSE, MAE, MAPE)

  • Configure Energy and CO2 emission tracking for the Green Network use case using CodeCarbon

    • we provide:

      • Training duration, RAM/CPU/GPU energy consumption, CO2 emissions

  • Upload the trained model along with its metrics (using ModelMetricsSdk)

• Step 3. Compile the pipeline code to generate a Kubeflow pipeline YAML file

...

• Set the TrainingJob name in lowercase

• Configure the Feature Filter

  • For the query to work correctly, use backticks(`) to specify a specific cell site for filtering (e.g., `Viavi.Cell.Name` == "S1S10/B13/C1C3")

• Refer to the image below

...

5. Result

• These logs can be reviewed through the logs of the Kubeflow pod generated during training execution, and the details that can be checked are as follows:

...

6. Load Model

 apiVersion: "serving.kserve.io/v1beta1"

kind: "InferenceService"

metadata:
   name: traffic-forecasting-model
  namespace: kserve-test
spec:

  predictor:

    tensorflow:
       storageUri: "http://tm.traininghost:32002/model/viavi_lstm_s1b13c3s10b13c3/1/Model.zip"
       runtimeVersion: "2.517.10-gpu"

      resources:
         requests:

          cpu: 0.1

          memory: 0.5Gi
 1Gi
          nvidia.com/gpu: 1
        limits:

          cpu: 0.1

          memory: 0.5Gi 1Gi
          nvidia.com/gpu: 1

kubectl create namespace kserve-test
kubectl apply -f deploy.yaml -n kserve-test

...

source predict.sh

...

7. Comparison

import json
import requests
import matplotlib.pyplot as plt

with open('input.json', 'r') as file:
    data = json.load(file)

instances = data['instances']
original_data = [instance[-1][0] for instance in instances[1:]]

print(f"Original: {original_data}")
print(f"Length: {len(original_data)}")

cluster_ip = "172.24.100.123"   # IP of where Kserve is deployed
ingress_port = "32667"          # Port of the Istio Ingress Gateway
url = f"http://{cluster_ip}:{ingress_port}/v1/models/traffic-forecasting-model:predict"

headers = {
    "Host": "traffic-forecasting-model.kserve-test.svc.cluster.local",
    "Content-Type": "application/json"
}

response = requests.post(url, headers=headers, json=data)

if response.status_code == 200:
    predictions = [pred[0] for pred in response.json()['predictions']]
else:
    print(f"Error: {response.status_code} - {response.text}")
    predictions = []

print(f"Predictions: {predictions}")

plt.figure(figsize=(35, 7))
plt.plot(original_data, label='Original Data', linestyle='-', marker='o')
if predictions:
    plt.plot(range(len(predictions)), predictions, label='Predictions', linestyle='--', marker='x')
plt.title('Comparison of Original Data and Predictions')
plt.xlabel('Index')
plt.ylabel('Value')
plt.legend()
plt.grid(True)
plt.show()

...

4. File list

• CellReports.csv
  The file contains traffic data from Viavi.
  

• View file
  name insert.py

  The file processes the dataset and inserts the data into InfluxDB.
  (Changed required: DATASET_PATH , INFLUXINFLUXDB_IP , INFLUXINFLUXDB_TOKEN)
  

• View file
  name pipeline.ipynb

  The file defines the model structure and training process.
  

• View file
  name deploy.yaml

  The yaml file is used for deploying model inference service.
  

• View file
  name predict.sh

  The script used for excuting the model prediction.
  

• View file
  name input.json

  The json file is used for Inference.

5. Example

• Input data

[
    [
        [
            0.8342111290054597
        ],
        [
            0.6010991527639614
        ],
        [
            0.6010991527639614
        ],
        [
            0.7556675063318373
        ],
        ...
    ]
]

• output data

"predictions": [
    [0.112432681], [0.0969874561], [0.100156009], [0.308645517],
    [0.303829], 0.41753903], [0.539769948], [[1.873328690.618172765],
    [0.634754062], [0.625748396], [0.530416906], [0.333626628],
    [0.208134115], [0.137386888], [0.113081336], [0.105671018],
    [0.14362359], [0.258447438], [0.55385375], [0.567666292],
    [0.680490911], [0.738528132], [10.91417336717839658], [0.348406911653536677],
    [0.607577682524426162], [0.453805923], [10.88934326268508255], [10.98845983170119792],
    [10.89937615115568757], [10.744735240889923126], [1.96166563], 
[2.02938437[0.0926849917],[0.143254936],
    [0.277066618], [0.629820824], [10.80582476710377932], [10.93089557728215456],
    [10.82921791707016706], [10.25594974638304889], [0.527941108], [0.942011893411735266],
    [10.86683917345081121], [0.336141407171168745], [10.7329919314139232], [0.8318588730903047472],
    [0.389349937231243849],[0.170835257], [0.482014239178088814], [10.81820059255523622],
    [10.92650104405331433], [10.6451056626908183], [0.738317609], [10.99126577797301173],
    [10.89580131575131953], [20.00482678416216075], [10.82429934296788305], [10.8287394203006953],
    [0.15064013], [10.78468418136724383], [10.86079311159134328], [0.384998083414751947],
    [10.85713887386262804], [10.56870174395934224], [0.432646126], [10.80646992598033369],
    [10.8532176657576084], [10.86798143664435327], [10.48708344585267484], [0.50265520840978539],
    [0.69283026528032586], [0.203143746], [10.35089064136435121], [0.512491345102216847],
    [0.109813243], [0.164212137], [0.438008606], [0.511521935]
]

Contributors

• Peter Moonki Hong - Samsung

• Taewan Kim - Samsung

• Corbin(Geon) Kim - Kyunghee Univ. MCL

• Sungjin Lee - Kyunghee Univ. MCL

• Hyuksun Kwon - Kyunghee Univ. MCL

• Hoseong Choi - Kyunghee Univ. MCL

Version History