How Artificial Neural Networks Can Help Data Center Management
By: Kayla Matthews on May 11, 2020
Artificial neural networks are computing algorithms loosely based on how animal brains learn. They get "smarter" by referring to examples and usually do not require task-specific programming to function.
Researchers are eager to see how neural networks could bring advantages to various industries and assist with needs often performed manually or based on guesswork. Data center management represents one application of neural networks already showing impressive potential. Here are four examples.
1. Keeping the Facility Cool
Maintaining adequate cooling in a facility is a constant concern for data center managers. They know how servers generate so much heat that they could quickly become too hot and shut down.
Google started applying neural networks to data center tech for cooling several years ago. By 2018, the company perfected its system enough that it wholly entrusted the artificial intelligence (AI) setup to make temperature adjustments without getting approval from humans. Before that point, the tool suggested changes, but people reviewed them before implementation.
Google's system has cloud-based AI at its core. It takes data from thousands of sensors every five minutes and gives the information to deep neural networks. Those networks predict the outcomes of billions of combinations of temperature-related actions. The goal is to figure out which one reduces energy usage the most while adhering to safety constraints.
This use of artificial neural networks results in an average energy savings of about 30%, and Google expects further improvements as it tweaks the technology over time. It's vital to clarify, too, that humans can exit the AI-driven mode at any time and make decisions without that help if desired. However, this is one example of how advanced technology can take care of an essential data center need.
2. Making Improved Power Predictions
When data center managers identify the most power-intensive applications and equipment, they can make responsive changes to cut the overall energy use. However, many of the established methods of doing that make calculations based on static relationships between power consumption and the hardware or applications responsible for it. Those techniques do not take in the real-life fluctuations occurring.
Researchers devised a better solution based on intensive power profiling and deep learning neural networks. The team began by evaluating different power series samples while utilizing a method to reduce noise interference.
They then developed two deep learning models to make power usage predictions. The results showed that the neural network approach offered up to a 79% error reduction in some cases, plus it was more accurate than other prediction approaches.
3. Forecasting Short-Term Traffic Loads
The best neural network algorithms are highly specific and well-trained to handle differences in the input. For example, the healthcare industry also uses this kind of AI to help with manual tasks, such as populating patient charts. Users have to train them first, however, so that the algorithms accurately recognize what someone says in a particular accent.
Thorough training helps the algorithms continue to work despite various minor changes, such as if a person's voice gets deeper or sounds slightly slurred because of tiredness. In the same way that algorithms learn to cope with those changes, some successfully predict the likelihood of fluctuations.
Going back to data center management, specifically, a research team became interested in applying artificial neural networks to predict short-term alterations in the amount of traffic handled by a facility. The group collected more than six million data points representing active users per second and used the neural networks to identify likely changes in traffic flow.
Similar applications could aid data center managers in avoiding the consequences of extreme traffic spikes, such as outages. Getting such short-term predictions could allow for better resource balancing to keep things running smoothly.
4. Predicting Failing Equipment
Another data center tech improvement involving these smart algorithms relates to using them for hard drive failure predictions. A team wanted to get a clear idea of the likelihood of failure for more than four million hard drives, and it needed two petabytes of system data to do it.
One of the near-term goals for this project was to make customers' server performance more reliable. Work is still underway, but the people carrying it out say that the results so far using the neural networks surpass the less-advanced methods used before. The team also hopes to get to the point of monitoring all kinds of connected equipment through a blend of traditional techniques and AI.
Fascinating Advancements in Data Center Management
These four examples show why people are getting more interested in applying artificial neural networks to manage data centers. Humans still play a crucial role in the respective tasks, but algorithms could pick up on things they miss.
About Kayla Matthews
Kayla Matthews writes about data centers and big data for several industry publications, including The Data Center Journal, Data Center Frontier and insideBIGDATA. To read more posts from Kayla, you can follower her personal tech blog at ProductivityBytes.com.