Flux Research Group University of Utah
No PDF available

WarmFlex: Coordinated Stabilization of Cold-Start Frequency and Idle Containers in Serverless Platforms

Ghazal Abdollahi, Hamid Asadi, and Robert Ricci

The 11th IEEE International Conference on Smart Cloud (SmartCloud) 2026

DOI: 10.1109/smartcloud69481.2026.00009

abstract
Cold starts persist in serverless platforms because providers must balance two competing goals: (i) keeping enough warm containers to reduce cold-start frequency, and (ii) releasing containers quickly to avoid prolonged idleness. Existing approaches often optimize one metric while treating the other as a constraint, leading to unstable operating behavior under heterogeneous workloads. We present WarmFlex, a function-centric keep-alive controller that stabilizes the joint operating regime of serverless execution by combining longevity-aware clustering derived from execution-time signatures with a fuzzy controller that adapts keep-alive duration using inter-arrival feedback. Trace-driven evaluation on the Azure Functions 2019 dataset shows that WarmFlex concentrates system behavior within a compact region of the cold-start/idle space, reducing the average container footprint relative to representative baselines while maintaining a desirable trade-off. The design is interpretable and deployable without the need for training pipelines.

areas
Cloud

BibTeX Hayagriva