Predictive Modelling and Cold Start Mitigation in Function-as-a-Service Using BiLSTM with Multi-Head Mechanism

Morla, Mohan Sai (2025) Predictive Modelling and Cold Start Mitigation in Function-as-a-Service Using BiLSTM with Multi-Head Mechanism. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (1MB) | Preview
Preview	PDF (Configuration Manual) Download (2MB) | Preview

Abstract

Serverless computing platforms, particularly Function-as-a-Service (FaaS) environments suffer from cold-start latency problems where function containers require initialisation time during the first invocation after idle periods by causing delays of hundreds of milliseconds to several seconds. This unpredictable latency severely affects real-time applications, machine learning inference services, and user experience, making intelligent cold-start mitigation essential for enterprise serverless adoption. This study developed three deep learning architectures LSTM, BiLSTM and Multi-Head mechanisms to forecast function invocation patterns. Using Microsoft Azure Functions dataset containing 1.98 million real-world traces these models generated predicted wait intervals between invocations. Apache OpenWhisk served as the FaaS platform where Apache JMeter continuously invoked custom workload through the invoker module with Docker containers monitored for warm pre-loaded in memory versus cold requiring initialisation states. The Multi-Head BiLSTM achieved an MSE of 10,958.96 and RMSE of 104.69, reducing cold starts from 28.5 percent to 9 percent across 50,000 invocations. This study advances serverless performance optimisation by introducing intelligent container lifecycle management on AWS EC2 infrastructure. Future work should address multi-platform validation, real-time adaptive learning and cost-benefit optimization for production deployments.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Kazmi, Aqeel UNSPECIFIED
Uncontrolled Keywords:	Cold Start; FaaS; BiLSTM; Multi-Head; Serverless Computing
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:	01 Sep 2026 09:20
Last Modified:	01 Sep 2026 09:20
URI:	https://norma.ncirl.ie/id/eprint/9722

Actions (login required)

View Item