Heterogeneous Hardware Support for Apiary

Abstract

Function-as-a-Service (FaaS) platforms are an appealing option for developers because they save time and money by eliminating the work spent managing application servers. Apiary is a novel FaaS platform which performs extremely well on datacentric tasks by tightly integrating computation and storage layers, eliminating the time spent transferring data between the two. Moreover, Apiary’s robust provenance system and straightforward programming model provide compelling reasons for developers to use it for both data-centric and compute-intensive tasks. In this paper, we detail a general architecture for using Apiary’s asynchronous programming model to implement compute-intensive tasks as external services. These external services are free to make usage of specialized hardware such as GPUs, which provide extremely good performance for many typical compute-intensive tasks such as machine learning inference.