SC20 Proceedings

The International Conference for High Performance Computing, Networking, Storage, and Analysis

Scheduling in Data Centers Running on Renewable Energy with Deep Reinforcement Learning

Workshop:2nd Workshop on Machine Learning for Computing Systems

Authors: Vanamala Venkataswamy (University of Virginia) and Andrew Grimshaw (University of Virginia, Lancium Inc)

Abstract: Most datacenters operate on electricity generated using non-renewable sources. Electricity cost is a significant part of the operational expense in datacenters. Higher electricity costs translate to endusers paying higher prices for cloud services or service providers seeing reduced profits. To reduce the cost to endusers and make higher profits, cloud service providers need to reduce operational expenses. Renewable energies are increasingly becoming a viable electricity source that dramatically lower electrical power costs and achieve dramatic reductions in climate impact. The green datacenters are colocated at the sources and powered by renewable energy. The green datacenters are not restricted to using a single source of renewable energy; instead, they utilize multiple energy sources.

Using renewable energy sources to power the datacenters has challenges. For instance, wind flow is not continuous and not uniform across regions, time of day, or seasons. The datacenters running on renewable energy sources need smart system-software that adapt to the power variability to ensure that cloud services are available even when there is transient power unavailability in these datacenters. Three issues that need addressing are 1) Meeting Service Level Objectives, 2) Resource Pool Management, and 3) Adapting to power variability.

Hand-engineering domain-specific heuristics-based schedulers to meet specific objective functions is time-consuming and expensive, and requires extensive tuning in this dynamic environment. We applied deep reinforcement learning (DRL) to automatically learn effective job scheduling policies while continuously adapting to the complex dynamic environment. The DRL-based scheduler's objective function is to maximize the total value from jobs.

Back to 2nd Workshop on Machine Learning for Computing Systems Archive Listing

Back to Full Workshop Archive Listing