Towards Quality of Experience Fairness for Serverless Functions

Serverless computing represents an innovative cloud computing model where the intricacies of server management are entirely concealed from users. It achieves this by running user-defined functions within transient containers. Cloud providers enhance resource efficiency by grouping containers from various users on a shared hardware infrastructure to offer Function as a Service (FaaS). However, this approach can result in significant disparity in the Quality of Experience (QoE) for end users. In this paper, we present Mable, a dynamic resource allocation technique for serverless FaaS platforms to address the challenges of achieving QoE fairness in the presence of shared resource contention and inter-function interference. Mable employs the Multi-Armed Bandit algorithm to systematically explore CPU resource config-urations that enhance QoE fairness among serverless functions. Extensive evaluations utilizing the FunctionBench benchmark and an open-source serverless computing platform, OpenFaaS, show that Mable significantly improves QoE fairness among a diverse mix of colocated functions under various workload con-ditions. Additionally, it reduces overall performance degradation of colocated functions. Mable outperforms the baseline (with no CPU regulation) and a heuristic-based CPU regulation technique by up to 89 % and 85 % respectively, in terms of eliminating unfair QoE. Notably, Mable manages to attain up to 16% reduction in CPU consumption while maintaining comparable throughput and average latency when compared to the baseline.