Computers play a vital role in the technology laden lives of today. As our demands for faster and more powerful computers continue to increase, so does the amount of energy that we use to power these machines. By 2025 data centers are forecast to be using 20% of all available electricity in the world. In 2014 a cloud provider used the equivalent energy consumption of approximately 366,000 U.S. households.
Additionally the growing complexity of computing systems, ranging from cell phones to supercomputers, is becoming difficult for human developers to manage on their own. As such, there is a need to develop resource management techniques for today’s complex high performing computer systems that minimize energy consumption without sacrificing compute speed.
Sandia is leveraging machine learning techniques, more specifically Reinforcement Learning (RL), to develop intelligent automated approaches to resource management. By using this technique we look to dynamically adjust job Powerstate (P-State) to save power without negatively impacting compute time. Researchers’ goal is to identify a dynamic P-state policy that can balance power and runtime across the execution of apps and multiple cores.
The growing complexity of computing systems and the heterogeneity of the hardware comprising these systems can potentially be managed by an intelligent automated mechanism saving energy while maintaining computing performance. Machine learning RL techniques could be successfully applied to other complex system resource management problems.