For some leading-edge thinkers and organizations, machine learning is starting to appear in practical applications. Although people have theorized about this subset of artificial intelligence for more than half a century, machine learning examples are finally making their way to the home territory of computer science -- the data center.
In a complicated environment, where there is data about data, why not use the power of computing to improve computing? It appears to be a concept whose time has come.
"Machine learning to manage the data center is finally starting to come to fruition," said Michele Goetz, principal analyst at Forrester Research in Cambridge, Mass.
Data centers: An ideal environment
Data centers are an ideal environment for machine learning since there is so much data available, said Christopher Yetman, COO at Vantage Data Centers, a data center services provider in Santa Clara, Calif. There is sensor data at the building and physical level, with servers and network devices generating vast amounts of information about their own operations. Humans have used that data in the past, but not to its fullest potential.
Christopher YetmanCOO at Vantage Data Centers, Santa Clara, Calif.
There are also sensors that generate data about air pressure, humidity, temperature and supply voltage and typically feed into a programmable logic controller, Yetman said. The data can then trigger specific actions at certain thresholds, like when a room begins to get too warm. However, it is mostly a reactive process, and data center management tools tend to look at some, but not all, of the activities going on within processors, storage arrays or networks.
Now, people build smart, physical sensors using inexpensive Raspberry Pi technology that can speak to Ethernet and controllers. In other words, the growing amount of data available will lead to more machine learning examples in the data center.
"If you feed a good machine learning system with lots of information, it will pick off things like usage patterns on the server and maybe correlate that to the weather," Yetman said. Machine learning systems will eventually note that rising humidity and wind from the west leads to higher temperatures and can change the behavior of the data center in anticipation, he said.
"To choose another example, a social media company might understand that its users are more active in the early evening; machine learning would notice this and might automatically bring more machines online in anticipation of need and to keep response times acceptable," he said. Machine learning could combine that awareness with an understanding of how much heat those machines generate and ramp up cooling ahead of time.
More machine learning examples in the data center
Machine learning takes in data without knowing the relationships it contains, but ultimately learns those connections through its process. For example, Virtual Power Systems (VPS), a provider of software-defined infrastructure for data centers in Santa Clara, Calif., has built intelligent power supplies and uses machine learning to determine how to manage them. The software can speak to the servers and the power systems, and determine the need for change.
"VPS can arrange to allow one rack to borrow power from another by having one run partially on batteries in order to divert power to another so that it can run a peak load," Yetman said.
A data center is a complex interaction of multiple mechanical, electrical and controls systems, said Jim Gao, a data center engineer and researcher at Google, from his study, "Machine Learning Applications for Data Center Optimization."
"The sheer number of possible operating configurations and nonlinear interdependencies make it difficult to understand and optimize energy efficiency," Gao said. In response to this problem, Google built a machine learning system to fine tune operations and save energy.
Similarly, researchers from Ireland invested in some machine learning technologies that directly apply to data center management and operations. "In analytics and cloud computing ... we are seeing interesting outputs for the data center at the intersection of these two technology domains in the form of autonomic systems that self-manage, self-configure, self-heal and self-protect," said Shane Nolan, senior vice president for technology, consumer and business services at Industrial Development Agency, a government agency that develops foreign investment, based in Ireland
Nolan added that hyperscales like Facebook, Apple, Google and Amazon are making Ireland one of the fastest growing data center locations in Europe. As a result, advances in these facilities and systems management in data centers are important.
Will machine learning take over?
Yetman's theory is that software kills hardware, in the sense that software capabilities have outstripped what could be accomplished using hardware alone. He believes the demise of fault-tolerant computer hardware vendors was due to the rise of technologies such as virtualization, which delivers similar advantages at lower costs. Similarly, he predicts, machine learning will kill software.
"Software is based on humans trying to anticipate every solution to a group of problems, but people never anticipate every situation and the software that they create is relatively inflexible," Yetman said. "That is why [machine learning] will kill software; it will learn the optimal way of doing things."
One of the classic machine learning examples of this capability is correlating multiple logs. Because the possible relationships between different logs are unknown, simple queries or preexisting report structures are unlikely to reveal anything new. However, allowing the machine to learn by looking for correlations can produce insights not achievable in any other way.
But despite the enthusiasm, machine learning examples for data center management are still emerging, with most being of the homegrown variety.
"Most of those developments are with the hyperscale organizations like Google," Yetman said. "For those kinds of organizations, the effort involved with developing a machine learning solution pays off faster, so they are willing to build it."
Examine the risks involved with machine learning
Machine learning examples continue to grow
An introduction to machine learning algorithms