Nothing stands still in the tech world. At least, not without being whizzed right by and forgotten in short order, that is. The data center slice of the tech world is no exception to that rule, and today in our blog we’re taking a look at how the owner of some of the world’s largest data centers is attempting to push the business forward by making it more efficient than ever before.
Who’s the data owner in question? We’ll give you a few hints: they’re big fans of bright colors; they’re pretty darn efficient at helping you find what you’re looking for (maybe they even helped you find this very blog post); they recently changed their logo so furtively that most of the world didn’t even notice; and they love a good mountain view. Yes, we’re talking about the titan of search, Google.
Google describes itself as being “obsessed with saving energy” (Us too!) and prides itself on its history of continually iterating on data center efficiency over the last decade to ensure that its data is stored, monitored, managed, accessed and scaled with maximum energy savings in mind. The search giant’s VP of Data Centers, Joe Kava, recently spoke at Data Centers Europe 2014 and the Google Blog released a whitepaper detailing its latest foray into saving energy at its centers: machine learning.
The Mountain View, CA-based company is hard at work constructing incredibly smart server farms capable of learning from their past performances and improving upon them in the future. This can mean only one thing: Skynet Google will eventually become self-aware and enslave mankind with an army of T1000s bearing a striking resemblance to a certain former California governor.
Well, no, actually it doesn’t mean that at all. What it means is that Google is designing a form of technology that could one day trickle down to other data centers and help your data be stored more efficiently than ever before.
The technology is the result of a “20 percent project.” What the heck does that mean, right? It’s pretty simple, actually. Google allows employees to find time to perform work that’s outside of the parameters of their official job descriptions. In this case, it came from 20 percent of the time of data center engineer Jim Gao.
A Boy Genius Among Geniuses
Gao is reportedly a smart guy – and let’s be honest here, he probably wouldn’t be working as a Google engineer if he wasn’t. Gao’s particularly intelligent, though, even for that subset of geniuses; his colleagues refer to him as “Boy Genius” because of his aptitude for breaking down and understanding huge sets of data. Mr. Smarty Pants himself was hard at work performing cooling analyses by using computational fluid dynamics that relies on monitoring data to form a three-dimensional airflow model within a server room – and who among us hasn’t been down that rabbit hole and back a million times, right?
In any case, Gao believed it possible to create a tracking model for a broader set of variables like climate conditions, IT load and cooling tower, heat exchange and water pump operations for keeping Google’s servers chilled out.
“One thing computers are good at is seeing the underlying story in the data,” Kava wrote in a blog post, “so Jim took the information we gather in the course of our daily operations and ran it through a model to help make sense of complex interactions that his team – being mere mortals – may not otherwise have noticed.
“After some trial and error, Jim’s models are now 99.6 percent accurate in predicting PUE [a measure of energy efficiency]. This means he can use the models to come up with new ways to squeeze more efficiency out of our operations. ”
The Thinking Man’s Data Center
So far it may sound like Gao was working on exactly what his job would require him to work on: data. And that’s true, but what the engineer did next was veer into new territory; he took a Standford University course on artificial intelligence for education in the realm of machine learning.
He needed to learn about neural networks, which mirror the functionality of the World’s Oldest Technology: the human brain. They permit computers to adapt to situations and learn tasks that they were never explicitly programmed to carry out. No, this is not science fiction, just information technology. Google’s most ubiquitous piece of tech (its search engine) actually relies on machine learning such as this. Gao used his learnings to create models that predict and improve data center performance.
“The model is nothing more than a series of differential calculus equations,” Kava explained. “But you need to understand the math. The model begins to learn about the interactions between these variables.”
He also noted that Gao’s designs are not unlike other existing examples of machine learning such as speech recognition software or a PC that analyzes significant amounts of data to identify patterns within it and learn from them.
There are many, many variables at play in a data center: user requests, outside air temperature, hardware capabilities, etc. Humans do their best to observe how they all work in concert with one another, but it’s hard. Really hard. Computers, on the other hand, boast stellar efficacy when it comes to seeing the aforementioned underlying story of data. Realizing this, Gao grabbed the info that Google’s data centers accumulate in the course of daily operations and put it through his model to sort out the intricate interactions.
Gao’s strikingly effective models came into play a few months ago when Google took some servers offline. Normally such a procedure means degrading the server farm’s energy efficiency. The model, however, lowered the impact of the PUE change by temporarily altering the cooling setup. These type of diminutive changes, when made on an ongoing basis, will result in enormous money and energy savings, according to Google.
“A typical large-scale DC [data center] generates millions of data points across thousands of sensors every day, yet this data is rarely used for applications other than monitoring purposes,” Gao wrote in his white paper. “Advances in processing power and monitoring capabilities create a large opportunity for machine learning to guide best practice and improve DC efficiency.”
Specifically, he found a way to make it feasible to meet target set points through a litany of hardware and software combinations. It would not be possible for humans to test the exhaustive list of potential combinations and find maximum efficiency levels. But for Gao’s model, it’s a piece of cake.
Leave Hollywood out of This One
Earlier in this post we made a requisite Terminator joke, but Kava assured everyone that Google’s centers are not on the path to true self-awareness anytime in the near future. The new machine learning tools aren’t here to run the show. They’ll be used precisely as their name applies: as tools.
“You still need humans to make good judgments about these things,” said Kava in his talk. “I still want our engineers to review the recommendations.”
Google foresees the neural networks being most useful in the building of future server farms as a forward-looking tool to test out design changes and innovations. And it would love to see the entire industry get in on it, which is why it shared Gao’s methodology in the whitepaper.
“This isn’t something that only Google or only Jim Gao can do,” said Kava. “I would love to see this type of analysis tool used more widely. I think the industry can benefit from it. It’s a great tool for being as efficient as possible.”
Image Source: Google