Clustering-Based Invocation Patterns Prediction in Serverless Computing Environments

This study presents a novel approach for predicting invocation patterns in serverless computing using clustering and classification techniques. Serverless computing introduces unique challenges for resource management due to its event-driven and highly dynamic nature, where invocation patterns can vary significantly based on unpredictable workloads. We address these challenges by employing KMeans clustering to segment invocation frequencies into distinct groups, each characterized by a unique signature defined by invocation count and execution time boundaries. These clusters are essential for understanding the functions' behavior and performance trends. They also serve as target variables for classification models, enabling the prediction of future invocation based on historical patterns. This study uses real-world traces from Azure Cloud's function apps that represent logical groups of a set of serverless functions sharing a runtime environment. We evaluate several classification models in our predictive framework, comparing their performance across multiple metrics. Also, we employ techniques such as Principal Component Analysis (PCA) for dimensionality reduction and sigmoid calibration to enhance prediction accuracy. This study offers valuable insights for improving decision-making in serverless management systems, enabling efficient allocation while reducing cost and operational overhead in cloud data centers.