Mitigating the cold start issue in Serverless using Reinforcement learning

Abstract

Serverless computing is a novel paradigm of cloud computing that allows cloud users to reduce underlying infrastructure down to the ”function” level of an application. Although, the function provides lightweight and efficient resource management, such control of resources comes at the expense of increased latency. The cold start problem might account for as much as 80% of the total latency, which creates the need for application optimization. This Master thesis proposes a two-tier approach for cold start mitigation utilizing Q-learning and K-means addressing latency, CPU, memory usage, and adherence to Service level objective. The Q-learning model takes into account the shortcomings of the existing approaches and is devised with adaptive to the latency level rewards due to the dynamic nature of the function invocations and uses multiple Q-tables to avoid long-size tables. The K-means clusters the invocations based on low, medium, and high latency. The method is trained with real-life function invocations from Huawei cloud and an evaluation dataset that was generated from function invocations in OpenFaaS. The method performed with 0.40 average reward per iteration during training and 0.46 and 0.47 during evaluation for one-by-one function invocation and parallel invocation respectively. In addition, the method proved cost efficiency by completing an iteration for around 5 seconds, required less than 8% CPU, and around 77 MB of memory for over 50 training iterations.