EcoTracker: An Application-Level Carbon Scheduling Framework for LLMs
Adomas Bagdonas, Georgi Dimitrov, Kristian Hristov, Stilyan Penchev, Daniel Rachev.
Group 19.
Paper.
Website.
Source code.
The rapid growth of Large Language Models has intensified concerns about the environmental cost of AI inference. Currently, developers using third-party LLM APIs lack plug-and-play solutions to actively mitigate their footprint. To address this, we developed EcoTracker, an open-source Python framework published on PyPI for application-level carbon-aware execution of delay-tolerant LLM workloads. EcoTracker combines forecast-guided temporal shifting with stochastic jitter to prevent artificial demand spikes. It also includes policy mechanisms such as dynamic model downgrading and carbon-budget enforcement. To ensure reproducible evaluation, we benchmarked the tool using a fixed carbon-intensity trace across 1,392 scheduling scenarios over 29 days. Relative to immediate execution, EcoTracker successfully reduced total emitted carbon by 9.99% and captured 76.9% of the savings achieved by an oracle policy with perfect hindsight, proving that meaningful emissions reductions can be achieved at the application layer.