Migrating Large Deep Learning Models to Serverless Architecture
Dheeraj Chahal, Ravi Ojha, Manju Ramesh, Rekha Singhal
Abstract
Serverless computing platform is emerging as a solution for event-driven artificial intelligence applications. Function-as-a-Service (FaaS) using serverless computing paradigm provides high performance and low cost solutions for deploying such applications on cloud while minimizing the operational logic. Using FaaS for efficient deployment of complex applications, such as natural language processing (NLP) and image processing, containing large deep learning models will be an advantage. However, constrained resources and stateless nature of FaaS offers numerous challenges while deploying such applications. In this work, we discuss the methodological suggestions and their implementation for deploying pre-trained large size machine learning and deep learning models on FaaS. We also evaluate the performance and deployment cost of an enterprise application, consisting of suite of deep vision preprocessing algorithms and models, on VM and FaaS platform. Our evaluation shows that migration from monolithic to FaaS platform significantly improves the performance of the application at a reduced cost.