Architectural Implications of Function-as-a-Service Computing

FaaS containerization brings up to 20× slowdown vs. native; cold start can exceed 10× a function's execution time

Featured image

Venue: MICRO 2019
Link: ACM DL

Topic: A comprehensive characterization of FaaS (Function-as-a-Service) workloads on OpenWhisk, identifying where time is spent, why cold starts are expensive, and what architectural bottlenecks exist.


Summary

FaaS containerization brings up to 20× slowdown compared to native execution. Cold-start latency can exceed 10× a short function’s execution time. This paper introduces a tooling framework (on OpenWhisk) to instrument and measure where time is spent across the FaaS stack, and analyzes over/under-invocation behaviors that affect latency.


Background

FaaS overhead sources

Invocation states and their effects

Over-invoked

Under-invoked

Balanced


Key Finding

Containerization overhead

Contribution


Insights


Meeting Notes

(to be filled)