A CROSS-CLOUD STUDY OF COLD START DETECTION IN AZURE AND GCP SERVERLESS ENVIRONMENTS

Cold start latency is a well‑known bottleneck in serverless computing, often causing noticeable delays when functions are invoked after periods of inactivity. In multi‑cloud environments, differences in architecture, resource management, and workload behaviour introduce additional complexity. An experimental study on predicting cold start events and comparing performance across two widely used serverless platforms: Microsoft Azure Functions and Google Cloud Platform Cloud Functions.

Large‑scale execution datasets are collected independently from both platforms and used to train and evaluate the models within their respective cloud environments to detect cold start events based on latency percentiles, request arrival patterns, memory usage, and temporal features to check cross cloud cold start detection the models were trained and tested separately on datasets from Azure-2019 and Google Cloud Functions-cold start dataset.

The results show that both the platform Azure and GCP have different features, so in the experiment, both datasets were trained and tested with their respective cloud environment, and the models used were Logistic Regression, XGBoost, and Random Forest classifiers. These models perform accurately in a cloud environment. But cross-cloud detection of cold start using machine learning performs poorly due to different architectures and features. The research suggests that cross-cloud cold start detection can be done by collecting real world dataset. In both the experiment, XGBoost and Random Forest out performs the Logistic Regression. Taken together, the findings emphasize the strength of tree‑based models in capturing complex, non‑linear performance dynamics that linear techniques are unable to represent effectively.