Machine learning behind the efficiency of Google’s data centers

Google’s cutting edge technology has brought some of the most innovative products to the market including the fastest search, browser and mail as well as a diverse range of online services. Efficient delivery of Google’s online services is made possible through its global network of data centers. Currently, there are sixteen of its data centers operational globally. This year Google has announced to invest around $19 Billions for expansion and development of new ones. Nine of its data centers are in the Americas and rest in Europe in Asia. Several more are planned to be developed this year in various parts of US.

Efficient operation of these data centers ensures consistent delivery of online services. Behind the efficiency of the Google data centers is Deep Mind, the world leader in research and application in the area of AI. Deep Mind was founded in 2010 and is currently a part of the Alphabet company which acquired it in 2014. It is constantly pushing the boundaries of AI and has brought breakthrough technologies like AlphaGo, the first computer program to have defeated a professional human Go player, Deep Q Network (DQN) and DNC or Differentiable Neural Computers which can learn from neural networks and store complex data as computers do.

DeepMind has also created several more breakthrough technologies for Google including the Wavenet a new deep neural network that produces better, more realistic sounding speech than existing techniques. In collaboration with Android, it has brought several new features that will deliver optimized performance to millions of Android Pi users. One of its biggest accomplishments is the machine learning system driving the efficiency of Google data centers. Google’s data centers are already highly optimized but Google consistently works to improve their efficiency as much as possible because of their central role in running a system supporting Google’s services at a global scale. It developed an AI powered recommendation system to improve the efficiency of Google’s data centers in 2016. Second phase of the same work began in 2018 whose purpose was to improve cooling of the data centers. The safety first AI system helps autonomously manage cooling of the data centers.

AI Powered Recommendation System

In 2016, Deep Mind and Google together developed an AI powered recommendation system which was founded on a basic idea that even minor improvements could help reduce energy consumption and CO₂ emissions. Google applied Deep Mind’s machine learning, leading to a reduction of 40% in the energy being consumed in the process of cooling. One of the most energy intensive processes in the operation of data-centers is cooling. These data-centers power several of Google’s services which people access worldwide including Gmail, Search and YouTube. In the process they produce a lot of heat. Pumps, chillers and cooling towers are used for cooling of data centers which can otherwise become inoperative due to the heat produced. However, achieving optimal performance may become difficult because of several factors. Data-centers and the environment interact in complex and nonlinear ways which cannot be deciphered using traditional formula based engineering and human intuition. Moreover, the different architecture and environment of each of these data centers and different operating scenarios make it difficult to adopt a custom solution.

However, Google used Deep Mind’s neural networks that were trained on various operating scenarios and parameters within the data centers for creating a more efficient model to understanding data center dynamics and optimize performance. A group of neural networks was trained on historical data collected using thousands of censors within the data centers. These neural networks were trained on future PUE or Power Usage Effectiveness since Google’s purpose was to achieve higher energy efficiency. PUE is the ratio of total building energy usage and IT energy Usage. Google also trained additional neural networks on future temperature and pressure over the next hour. This machine learning system enabled a 40 percent reduction in the amount of energy consumed in cooling.

Safety First Artificial Intelligence

Google took its machine learning system developed with the help of Deep mind (AI Powered recommendation system) to the next level in 2018. Now, it was time to automate the system to achieve higher efficiency. The recommendation system is replaced by an AI powered system that controls data centre cooling directly. The cloud based AI system is working across several data centers and driving energy savings and higher efficiency. A neural network fed by the cloud based AI system ever five minutes predicts how various actions can affect energy savings and which one will minimize energy consumption. There are several safety constraints to satisfy in the process too. This set of actions is sent to the data center where the local control system verifies them before implementation. High confidence actions are taken into consideration and the low confidence actions are eliminated. The operators of these data centers are always in control and verify the optimal actions against an internal set of constraints. Moreover, the operators can exit AI control mode any time. The system has been delivering consistent energy savings since its implementation and continues to grow better with time. In future, Google expects the implementation of this system in the other industries too could help grow productivity and help combat environmental threats at a larger scale.