FMI: The FaaS Message Interface

Abstract

Serverless platforms provide massive parallelism with very high elasticity and fine-grained billing. Because of these properties, they are increasingly used for stateful, distributed jobs at large scales. However, a major limitation of the commonly used platforms is communication: Individual functions cannot communicate directly and using external storage or databases for ephemeral data can be slow and expensive. We present FMI, the FaaS Message Interface, to overcome this limitation. FMI is an easy-to-use, high-performance framework for general-purpose communication in Function as a Service platforms. It supports different communication channels (including direct communication with our TCP NAT hole punching system), a model-driven channel selection according to performance or cost, and provides optimized collective implementations that exploit characteristics of the different channels. In our experiments, FMI can speed up communication for a distributed machine learning job by up to 1,200x, while reducing cost at the same time by factors of up to 365. It provides a simple interface and can be integrated into existing codebases with a few minor changes.