Function-as-a-Service Application Service Composition
Mohammadbagher Fotouhi, Derek Chen, Wes Lloyd · 2019
Serverless computing platforms provide Function-as-a-Service (FaaS) to end users for hosting individual functions known as microservices. In this paper, we describe the deployment of a Natural Language Processing (NLP) application using AWS Lambda. We investigate and study the performance and memory implications of two alternate service compositions. First, we evaluate a switchboard architecture, where a single Lambda deployment package aggregates all of the NLP application functions together into a single package. Second, we consider a service isolation architecture where each NLP function is deployed as a separate FaaS function decomposing the application to run across separate runtime containers. We compared the average runtime and processing throughput of these compositions using different pre-trained network weights to initialize our neural networks to perform inference. Additionally, we varied the workload dataset sizes to evaluate implications of inferencing throughput for our NLP application deployed to a FaaS platform. We found our switchboard composition, that shares FaaS runtime containers for all application tasks, produced a 14.75% runtime performance improvement, and also a 17.3% improvement in NLP processing throughput (samples/second). These results demonstrate the potential for careful application service compositions to provide notable performance improvements and ultimately cost savings for application deployments to serverless FaaS platforms.