MASTER: Machine Learning-Based Cold Start Latency Prediction Framework in Serverless Edge Computing Environments for Industry 4.0

The integration of serverless edge computing and the Industrial Internet of Things (IIoT) has the potential to optimize industrial production. However, cold start latency is one of the main challenges in this area, resulting in resource waste. To address this issue, we propose a new machine learning-based resource management framework called MASTER which utilizes an extreme gradient boosting (XGBoost) model to predict the cold start latency for Industry 4.0 applications for performance optimization. Furthermore, we created a new cold start dataset using an IIoT scenario (i.e. predictive maintenance) to validate the proposed MASTER framework in serverless edge computing environments. We have evaluated the performance of the MASTER framework using a real-world serverless platform, Google Cloud Platform for single-step prediction (SSP) and multiple-step prediction (MSP) operations and compared it with existing frameworks that used deep deterministic policy gradient (DDPG) and long short-term memory (LSTM) models. The experimental results show that the XGBoost-based resource management framework is the most successful model in predicting cold start with mean absolute percentage error (MAPE) values of 0.23 in SSP and 0.12 in MSP. It has been also identified that the Linear Regression model (utilized in the MASTER framework) has the least computational time (0.03 seconds) as compared to other deep learning and machine learning models considered in this work. Finally, we compare the energy consumption and $\text{CO}_{2}$ emissions of all models to emphasize resource awareness.