Mitigating Cold Start Problem in Serverless Computing: A Reinforcement Learning Approach
Parichehr Vahidinia, Bahar Farahani, Fereidoon Shams Aliee
Abstract
Serverless computing has revolutionized the world of cloud-based and event-driven applications with the introduction of Function as a Service (FaaS) as the latest cloud computing model. This computational model increases the level of abstraction from the infrastructure and breaks the program into small units called functions. Thus, it brings benefits, such as ease of development, saving resources, and reducing product launch time for enterprises and developers. Thanks to the scale-to-zero feature of this computational model, idle functions with no traffic will be depreciated from memory. However, this cost-saving approach adversely impacts delay leading to the cold start problem. Unfortunately, the existing solutions to alleviate the cold start delay are not resource efficient as they follow a fixed policy over time. Thereby, this article proposes a novel two-layer adaptive approach to tackle this issue. The first layer utilizes a holistic reinforcement learning algorithm to discover the function invocation patterns over time for determining the best time to keep the containers warm. The second layer is designed based on a long short-term memory (LSTM) to predict the function invocation times in the future to determine the required prewarmed containers. The experimental results on the Openwhisk platform show that the proposed approach reduces the memory consumption by 12.73% and improves the execution invocations on prewarmed containers by 22.65% compared to the Openwhisk platform.