Peeling Back the Carbon Curtain: Carbon Optimization Challenges in Cloud Computing

The increasing carbon emissions from cloud computing requires new methods to reduce its environmental impact. We explore extending data center server lifetimes to reduce embodied carbon emissions (from hardware manufacturing), rather than operational (from running hardware). Our experiments are the first to analyze a data center application's end-to-end performance on different server generations, to reveal that older hardware can preserve performance in certain conditions (e.g., low load). Our observations show the need for a carbon-aware data center scheduler that schedules on older hardware when suitable. However, quantifying such a scheduler's carbon savings is challenging today due to the lack of practical carbon measurement metrics/tools. We identify gaps in current methods for measuring operational and embodied carbon and call upon the broader systems research community to take action and conduct research that can pave the way for future carbon footprint analysis in systems.