An Adaptive ML-Driven Layer-wise Caching and Pre-warming Approach for Cold Start Mitigation in Serverless Frameworks

Udumula, Pavan Kumar Reddy (2025) An Adaptive ML-Driven Layer-wise Caching and Pre-warming Approach for Cold Start Mitigation in Serverless Frameworks. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (700kB) | Preview
Preview	PDF (Configuration Manual) Download (492kB) | Preview

Abstract

Applications that rely on serverless computing with AWS Lambda can experience delays since it takes time for processes to start which can slow down time-sensitive tasks. The current strategies do not work well in the cloud because static caching does not respond to changing demands and large predictive pre-warming is expensive and still fails when there is a sudden increase in traffic. This research offers an adaptive framework that merges intelligent dependency caching with pre-emptive prediction to reduce the cold start challenges in AWS Lambda environments. The framework utilizes a hybrid LSTM-XGBoost model to predict the invocation patterns of functions, where LSTM networks capture the temporal relationships among the different functions and XGBoost models capture feature interactions across system metrics. The hybrid prediction system offers cold start probability estimation with 67.3 percent combined accuracy and the confidence levels ranging between 45.1 to 46.6 percent at different load settings. The performance assessment shows significant improvements in latency with a reduction in response time during optimal warm execution conditions to about 92.5- 94.0ms. The three-tier caching system configuration (hot, warm, cold) with dynamic resource allocation delivers up to 20 percent cost savings through pre-warming strategies. The analysis provides validation to the increased performance optimization in serverless computing by exhibiting the quantifiable benefits of optimizing the cold start frequency, execution time, memory usage, and costs-per-million invocation.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Emani, Sai UNSPECIFIED
Subjects:	T Technology > T Technology (General) > Information Technology > Cloud computing Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Machine learning
Divisions:	School of Computing > Master of Science in Cloud Computing
Depositing User:	Ciara O'Brien
Date Deposited:	31 Mar 2026 10:45
Last Modified:	31 Mar 2026 10:45
URI:	https://norma.ncirl.ie/id/eprint/9273

Actions (login required)

View Item