Improve Cold Start Latency for Stateful Serverless Functions Using Pooled Resource Injection

Doolin, Tony (2025) Improve Cold Start Latency for Stateful Serverless Functions Using Pooled Resource Injection. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (739kB) | Preview
Preview	PDF (Configuration Manual) Download (1MB) | Preview

Abstract

Serverless computing allows applications to be easily deployed and takes away a lot of the complexity of scalability and maintainability. However cold start latency issues limit its adoption for stateful applications like Online Transaction Processing (OLTP) systems. This problem is compounded by the additional cost of establishing network connections to backed database systems for each serverless function invocation, as these functions cannot maintain state between invocations. This research investigates a novel approach to reducing cold start latency by injecting pre-established TCP connections into serverless containers during startup. A Connection Pool Manager was implemented to maintain active network connections and distribute them to serverless functions via Inter-Process Communication using file descriptor passing. Experimental evaluation using a serverless function deployed on a Docker container in the cloud compared direct connection establishment against connection injection across 1000 test iterations for varying connection establishment speeds. Results show that connection injection achieves connection establishment times 3.6 times faster than direct connections (260µs vs 951µs average). It also has reduced variance and improved predictability. Overall container lifecycle times improved by 4.4%, with 99th percentile response times decreasing from 514ms to 439ms. However, the research also encountered practical limitations regarding TLS connection injection and Kubernetes network namespace restrictions which limits its use in the real world. Notwithstanding this, this work provides valuable insights for improving serverless performance in stateful use cases and offers a possible approach for migrating OLTP workloads to serverless platforms.

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:	20 Mar 2026 15:35
Last Modified:	20 Mar 2026 15:35
URI:	https://norma.ncirl.ie/id/eprint/9212

Actions (login required)

View Item