Unipool: Efficient Management of Pre-Provisioned Resources for Container Platforms
Abstract
Cloud platforms reduce deployment latency by pre-initializing resources, but fragmented resource management across cloud layers often leads to substantial over-provisioning and increased cost. Through production trace analysis and replay using state-of-the-practice resource management approaches, we show that independently managing resources at each layer results in a poor trade-off between resource waste and deployment latency. We present Unipool , the first unified platform that exposes the state of pre-provisioned resources across layers, enabling more accurate resource pool sizing and dynamic multiplexing. Guided by insights from production traces, Unipool leverages cloud-native workload patterns to safely consolidate and oversubscribe pre-provisioned resources, significantly reducing platform cost while preserving quality of service. We implement Unipool on Kubernetes and Kata Containers, and evaluate it using both a large-scale event-driven simulator and a testbed of 25 virtual machines. Across mixed Function-as-a-Service and Container-as-a-Service production traces, we reduce resource waste by 92% while lowering container deployment latency by 21% compared with state-of-the-art baselines.