EcoFreq: Compute with Cheaper, Cleaner Energy via Carbon-aware Power Scaling

High-performance computing (HPC) constitutes an energy-hungry endeavor, and any efficiency gains via hardware and/or software advances are (over-)compensated by increased consumption (rebound effect). A large proportion of electricity is still generated by burning fossil fuels (mostly coal and gas), which is the cause of climate change, and has other potentially dangerous ecological and political consequences. Moving to a 100% renewable grid requires a plethora of solutions for electricity generation, distribution, storage, and consumption. In particular, dynamic load shifting can better align electricity consumption with the variable availability of solar and wind power. We introduce EcoFreq, a tool for dynamic power scaling on CPUs and GPUs that allows to maximize the usage of renewable, low-carbon energy. We benchmark EcoFreq via 14 HPC workloads using historical electricity market data from Germany, the UK, and the US. We show that carbon-aware power scaling leads to an over-proportional reduction in both, CO2 emissions and energy costs (e.g., 15% to 19% savings with an induced throughput decrease of only 10%). Furthermore, we observe that simple, static power capping at 70% - 80% results in considerably improved energy efficiency and we hence recommended it as default setting. EcoFreq is freely available at: https://github.com/amkozlov/eco-freq.

---

Contributors:

• Alexandros Stamatakis