The tech industry giant Google made headlines last year when the company announced it would be handing control of cooling systems in several of its hyperscale data centers to an artificial intelligence (AI) algorithm developed by DeepMind. The AI program underwent testing for several years, making recommendations to data center managers that could either be implemented or ignored. Either way, the decision to hand direct control over to the algorithm represents a major step in the development of AI.
Google’s ongoing experiments with AI yielded significant results. Over the course of testing, the AI’s optimization recommendations resulted in a 40 percent reduction in the amount of energy used for cooling. The algorithm’s performance demonstrated the energy savings potential of deploying an autonomous control system at scale. Although handing control of a facility’s cooling infrastructure over to AI may give some data center managers pause, the system still incorporates human oversight. The algorithm can function independently, but someone is still watching what it does and can override its decisions if the program behaves erratically or takes too many risks.
In this respect, it may be more accurate to think of an AI system as engaging in autonomous collaboration rather than autonomous control. Firm safety measures prevent the program from doing something that might damage cooling systems, and every decision it makes can be documented and tracked, often in real time. Since those decisions are backed up by data, it’s possible to review facility performance and track the algorithm’s confidence level over time.
While using AI to manage data center operations is generally described as a form of automation, the truth is a little more complex. Automation isn’t quite the same thing as autonomy. When systems are automated, they are given very specific protocols and preprogrammed procedures to follow. If an automated system encounters a situation it doesn’t recognize or have predetermined responses for, it won’t be able to respond at all.
Autonomous AI programs are different. They utilize machine learning and data center analytics to continuously gather data that allows them to respond to new situations. To give a crude example, a climate control system that turns on or off and adjusts temperature according to a set of well-established criteria or deterministic logic paths (such as specific “if, then” conditions) is automated. A system that knows how to adjust the temperature based on historical usage patterns (such as people making changes manually) is autonomous because it learned how to establish conditional responses on its own.
This is especially important in a data center environment. These facilities are extremely complex, with countless interactions between computing equipment, cloud computing workloads, cooling and power infrastructure, and human operators happening every moment of every day. Furthermore, data centers are usually purpose-built, meaning that each facility is a unique environment and an automated system designed for one may work for another.
Autonomous AI programs use machine learning and data center analytics to adapt to new situations and implement strategies to manage infrastructure more efficiently. The modern data center is far too complex to preprogram every possible response to every possible situation. Not only are there too many parameters to take into account, but there are also many scenarios that human programmers or even computer simulations might not be able to predict. Machine learning protocols allow AI programs to react to new situations by analyzing past events and making a probabilistic determination of how to best respond.
Handing cooling system management in a data center over to AI algorithms makes sense because cooling infrastructure is difficult for humans to manage effectively. Data center cooling needs are often driven by the power consumption of computing equipment, which is itself based upon usage patterns. In a complex data center environment with hundreds of servers operating simultaneously, there are far too many factors in play for human beings to account for in real time. A sophisticated AI system powered by machine learning and data center analytics not only analyzes what’s happening on the data floor in that precise moment, but also compares that analysis to previous situations and determines what adjustments need to be made to achieve the optimal outcome.
While AI may eventually take over other areas of data center operations, cooling is an ideal place to begin implementing this new technology because the potential efficiency gains are so significant. Operating data center cooling infrastructure more efficiently not only means less power consumption on a day-to-day basis, but it also means less wear and tear on cooling equipment and reduced chances of computing equipment being damaged due to poor environmental control (such as a server overheating or suffering moisture damage from condensation).
As artificial intelligence technology continues to develop in leaps and bounds, additional areas of data center operations will likely be handed over to autonomous systems that operate in tandem with human oversight. Given the efficiency gains Google has already experienced with this technology and similar data center analytics, the future of energy efficient data centers that capitalize on sustainable energy sources may well depend upon machine learning algorithms.