Designing Cloud Servers for Lower Carbon
- Jaylen Wang ,
- Daniel S. Berger ,
- Fiodar Kazhamiaka ,
- Celine Irvene ,
- Chaojie Zhang ,
- Esha Choukse ,
- Kali Frost ,
- Rodrigo Fonseca ,
- Brijesh Warrier ,
- Chetan Bansal ,
- Jonathan Stern ,
- Ricardo Bianchini ,
- Akshitha Sriraman
ISCA |
Major cloud providers intend to reduce carbon emissions by 2030, which requires effective interventions with short deployment timelines. We find that designing carbon-efficient compute server SKUs, or GreenSKUs, is a promising avenue as compute servers cause the majority of cloud emissions. However, designing GreenSKUs has several adoption challenges.
To address GreenSKU design challenges for the first time, we develop a systematic methodology and associated framework, GSF, that helps cloud providers make informed GreenSKU design and deployment decisions. GSF enables designing GreenSKU servers and evaluating their end-to-end performance and carbon impacts. GSF accounts for tradeoffs between different emission types and application performance requirements. We apply GSF within a leading cloud provider’s production constraints to make a systematic case for designing and deploying GreenSKUs.
We build a new GreenSKU and use GSF to show that it reduces carbon emissions per customer core by 29% compared to currently-deployed cloud servers. When deploying GreenSKUs in a way that meets applications’ performance requirements, we reduce emissions by 16%. When incorporating overall data center overheads, GreenSKU reduces cloud emissions by 9%.