Enabling Sub-second QoS-Aware Scheduling for Dynamic Serverless Workloads
摘要
Serverless computing platforms face significant challenges in fulfilling the Quality-of-Service (QoS) demands of serverless functions. Existing resource management strategies rely on time-consuming online fine-tuning processes and are limited by the static nature of container resource configurations. Moreover, the cold start problem leads to unpredictable performance degradation under dynamic workloads. To address these challenges, we propose a novel system that integrates Process-Level Checkpoint/Restore (PLCR) techniques with a Real-time Workload-and-QoS-Aware Scheduling algorithm. PLCR enables rapid switching of serverless functions within the lifespan of a container instance, effectively mobilizing idle containers to assist functions at risk of QoS degradation while significantly alleviating cold start delays. The scheduling algorithm dynamically adapts to workload fluctuations and workflow changes to maintain sub-second responsiveness and proactively ensure QoS compliance. Experimental results demonstrate significant improvements in the system’s ability to handle dynamic workload variations, with our PLCR mechanism achieving 2.2 \(\times \) –28.1 \(\times \) faster function switching compared to cold starts, reducing QoS violation rates by up to 13.5% across different functions, and outperforming the state-of-the-art Aquatope system by over 15% in QoS compliance during the critical initial deployment phase.