Leveraging static analysis for cost-aware serverless scheduling policies

Mainstream serverless platforms follow opinionated, hardcoded scheduling policies to allocate functions on the available workers. Such policies may decrease the performance of the application due to locality issues (e.g., functions executed on workers far from the data they use). APP is a platform-agnostic declarative language that mitigates these problems by allowing serverless platforms to support multiple, per-function, scheduling logics. However, defining the “right” scheduling policy in APP is far from trivial, often requiring rounds of refinement involving knowledge of the underlying infrastructure, guesswork, and empirical testing. We propose a framework that lightens the burden on the shoulders of users by deriving cost information from the functions, via static analysis, into a cost-aware variant of APP that we call cAPP. We present a prototype of such framework, where we extract cost equations from functions’ code, synthesise cost expressions through off-the-shelf solvers, and implement cAPP to support the specification and execution of cost-aware allocation policies.