Evergreen: Comprehensive Carbon Model for Performance-Emission Tradeoffs

The pervasive proliferation of computing infrastructure in recent decades has led to an increase in the fraction of worldwide energy and greenhouse gas (GHG) emissions associated with computing. A further steep increase is projected for the future, especially in light of the computational demands of large language models. While computing research has traditionally focused on performance, power, and area optimization, climate change concerns demand that the carbon footprint (CF) associated with computation be raised to a primary design parameter on par with power, performance, and area.To address this need, we propose Evergreen, a framework that augments computing tradeoffs with emission metrics. Evergreen offers the first complete carbon model for computing emission analysis, taking into account all major sources contributing to the carbon emissions associated with the execution of an application in a data center. The Evergreen carbon model considers both operational and embodied emissions associated with the computing hardware executing the application, the data transmission infrastructure, the renewable energy sources, and the battery energy storage systems. Using this model, our framework provides both an emissions predictor and a user-driven, emission-aware scheduler. The Evergreen scheduler allows users to select the data center to execute their workloads based on emissions and performance constraints associated with their execution in a cloud environment. Our evaluation shows how, for a given workload, Evergreen can provide multiple Pareto-optimal solutions in the performance-emission domain. Depending on the power source and data center values, Evergreen can identify carbon-optimal solutions that offer emission reduction with minimal latency overhead.