Friday January 17, 2025
- Share
- Share on Facebook
- Share on X
- Share on LinkedIn
Analyzing Renewable Energy and Power Capping in Data Centers
Abstract :
"Reducing greenhouse gas (GHG) emissions from Information and Communication Technology (ICT) has become a hot topic since the Paris Agreement. Data centers are one of the most impactful ICT energy consumers since they are built to run 24 hours / 7 days. This consumption has a big impact on GHG emissions when the power production comes from polluting sources, such as gas, coal, and oil. Several works were proposed to improve the energy spent by the processors. However, some authors pointed out a reduction in these improvements. This presentation focuses on two other ways to reduce emissions: Migration to renewable energy and limiting power usage (power capping). The presentation shows three contributions.
The first contribution is related to a renewable-only centralized data center. It focuses on online decisions, such as scheduling jobs and energy usage adjustments. A heuristic is proposed to mix production and demand predictions with online adaptations. This heuristic seeks to reduce the number of killed jobs and wasted energy. The second contribution considers renewable energy and power capping. In this contribution, we proposed a geo-distributed data center marketplace only supplied by renewable energies. The geo-distributed data centers compete for the users' jobs. The winner data center is defined using a reverse auction process. Our results show that a geo-distributed data center architecture in a cooperative marketplace is more efficient, earns more money, wastes less energy, and provides higher QoS than other approaches. Finally, we focus on power capping for a centralized HPC data center, presenting a lightweight scheduling algorithm. The algorithm's first step is predicting the power consumption, using the forecast results to evaluate if the system stays under the power capping. We show, using simulation, that a lightweight history-based prediction method can provide accurate enough power prediction to improve the energy management of a large-scale supercomputer compared to energy-unaware scheduling algorithms. In addition, we propose a knapsack algorithm for determining a good trade-off between performance and power limits.
To conclude, the presentation goes from mathematical models and scheduling algorithms to simulated experiments using traces and data from different sources."
Date et lieu
Friday, 17 January at 14:00
IMAG building, room 406
Organized by
Danilo CARASTAN-SANTOS
DATAMOVE Team
Speaker
Igor NARDIN
IRIT-Toulouse
- Share
- Share on Facebook
- Share on X
- Share on LinkedIn