μ: A bolt-on approach for faster and cheaper serverless computing

Serverless computing shifts the responsibility of managing and configuring the cloud infrastructure from the user to the provider. However, in the FaaS model, users are still faced with non-trivial resource allocation and management choices, in particular with respect to the way that intra- and inter-function parallelism is managed. As a consequence, deploying a cloud function with an incorrect configuration for either of these two dimensions of parallelism can lead to an unnecessary increase in execution time, cost, or both.

In this paper, we call the attention to the fact that this is one of the final hurdles to realizing the vision behind server-less computing. To overcome it without requiring changes to the existing cloud infrastructure, we propose the design of μ, a system that transparently manages the balance between intra- and inter-function parallelism on behalf of the user. μ is designed as a shim layer to be mostly transparent to the application developer and fully compatible with today's cloud services. Our design allows for running multiple tasks in a single function, using the extra resources that the user is already paying for. This allows for maximizing resource utilization, and leads to cost reductions and performance improvements, without the requiring the user to reason about resource allocation.