Cold Start Prediction and Provisioning Optimization in Serverless Computing Using Deep Learning

Abstract

Serverless computing has emerged as a significant framework for application development, offering benefits such as simplified deployment and enhanced developer productivity. Serverless designs accelerate app development, but user experience and performance are put at risk through the delay during a cold start. In this paper, an optimized concurrent provisioning methodology for the AWS Lambda environment has been proposed along with a cold start prediction technique based on deep learning. Employing historical data and real-time features like timestamp, invoke frequency, cold start indicator, duration, previous cold starts, event type, historical cold starts, time since the last cold start, and consecutive cold starts, an attention-based bi-directional gated recurrent unit (ABiGRU) is used to predict the cold start occurrences with high precision. In the AWS Lambda environment, the proposed DL model was able to predict the cold start likelihood for incoming Lambda invocations very accurately. In addition, the performance of the ABiGRU model is enhanced by hyperparameter tuning using the RMSProp optimizer. The findings of the experiments establish the proposed DL model to perform in the reduction of cold starts compared to the existing approach. Further, the ODL-CSP technique achieves an accuracy of 90.36%, a precision of 91.87%, a recall of 90.42%, an F1_score of 90.28%, and an MCC of 82.28% when applied to the testing dataset. Additionally, the proposed paradigm optimizes Lambdas using provisioned concurrency, similar to a function warmer. The proposed DL paradigm will eliminate cold start times by early deployment of Lambdas so that the ice age of the serverless architecture is eliminated.