Green Functions as a Service

How do we design distributed computing infrastructure with carbon as the first-order system-wide objective? This is the high level goal of this project.

``Green Functions as a Service'' is an abstraction and framework for providing managed sustainability, similar to how conventional FaaS provides managed scalability and resource allocation. It will provide a common platform for carbon and latency optimized execution of a wide range of latency-sensitive applications (such as ML inference, data analytics, etc.) on heterogeneous edge-cloud infrastructure.

Our approach will be grounded in fundamental principles of sustainability such as demand response, carbon pricing, and eco-feedback, and use modern AI techniques such as surrogate models for carbon modeling and optimization. We will extend serverless functions with new mechanisms for carbon-efficient execution, accurate carbon footprint modeling, and distributed resource management algorithms for reducing carbon emissions.

To reduce both the carbon footprint and latency, we envision a bottom-up approach:

1. Sound and complete energy measurement and attribution techniques. We will develop the notion of energy surrogates—statistical models providing energy and carbon footprints of functions. We will develop new statistical energy profiling and fair-division methods to provide fine-grained, accurate, and consistent energy footprints and carbon pricing.
2. The second thrust will develop new mechanisms for spatio-temporal demand repsonse, such as speculative function execution and new queueing policies. We introduce the notion of polymorphic functions for making energy-efficient use of heterogeneous hardware.
3. Dynamic routing and load balancing of functions in geo-distributed edge-clouds with different carbon and latency tradeoffs. We propose a new resoure allocation framework using carbon credits, and state-aware function migration mechanisms and algorithms which will incorporate statistical learning.

• Iluvatar serverless control plane: code