Bringing AI Under Control: The development of AI in data centers

By Claire Swedberg | Oct 15, 2019
shutterstock / nadya c

Data centers are part of everyday life—from enabling email storage to internet searches and navigation apps—and because of the growing demand by users, data center operators are feeling pressure to reduce expenses and improve power-usage effectiveness. Today, artificial intelligence, often using sensor data, helps with power management, including tracking, predicting and managing temperatures around each server; controlling cooling; and communicating with intelligent LED lighting systems.

While companies offer a variety of power management systems, there’s no single solution. That is in part because of the unique demands on each center, said Dave Schaible, Nlyte Software’s senior director. Data centers must balance operating expenses with maintaining the highest level of data availability.

Companies, such as Nlyte, provide power-monitoring solutions and collect data center power-use information and then correlate that data to the locations, rooms, server racks, IT devices and applications. This gives energy managers a view of how much power is being used throughout the data center and across a portfolio in real time. Past energy-use data is also available.

Traditionally, energy management systems have been run in isolated applications with limited access and generally focused on a single site. That required compiling data manually for comparisons.

“Nlyte’s solution includes role-based access control, so a broader audience can consume the data directly,” Schaible said.

Consolidating power-management data in one place makes comparing and planning efficient.Some utility providers and governments are offering rebate programs to upgrade older, inefficient infrastructure.

“Part of these programs requires monitoring to be in place to document the changes,” he said.

Machine learning and AI are the next evolution for managing power and energy in data centers, Schaible said. Since there are so many streams of information coming from both facilities and technology, the human operator can’t process it all.

“Leveraging things like IBM Watson will enable the data centers to adapt rapidly to changes,” Schaible said.

Electrical contractors would be best recommended to partner with a data center monitoring company at the start of a project, “to gain their input on the best devices for the application,” Schaible said. “There are so many options on the market today that there is no single device that can be used in all cases.”

Intelligent power

Since data centers are coming under growing scrutiny over the amount of energy they consume, there is a relentless push to improve the efficiency of the power use, said Marc Cram, director of OEM and global accounts, Server Technology Inc., Reno, Nev. Knowing the quantity of power used and how it was deployed (for instance, cooling, networking, computing, storage or lighting) are just two of the challenges faced by today’s data centers, he said.

Coordination between the operating system, the connecting “orchestration” layer and the physical hardware of the data center can enable the most time-critical applications to run first, receive power to run full throttle and prioritize the input and output from the data center. For instance, temperature can be measured at the room, rack and individual device level. Cooling can then address temperature excursions at the row, rack or individual IT device level, which allows the IT hardware to run at its most efficient and hottest temperature, Cram said.

AI is used to optimize the settings of the cooling systems of the data center, the networks and the servers themselves in order to wring the highest efficiency possible from each. AI may open and close dampers, turn off lighting and turn off servers.

A common mistake with temperature is relying on the room condition and not the rack, for instance. Temperature should be measured at the front and back of the rack with six sensors, mounted at the top, middle and bottom for a more accurate depiction, said Paul Mott, global technical product manager at Raritan. Raritan is a Legrand brand that sells data center power distribution units (PDUs). Humidity can be as important as temperature.

“Relative humidity should be between 40 and 60%. If humidity is too low, this will result in the buildup of electrostatic discharge, and if too high, corrosion will begin and slowly damage equipment,” he said.

Some data centers are using AI and machine learning technologies to analyze collected power and sensor data to help automate and optimize IT infrastructures. These smart tools will continue to advance and evolve to meet new data center challenges, Mott said.

Server Technology makes intelligent power products in a “configure to order” operation, “meaning we deliver the functionality and physical arrangement of outlets, cordage and input plugs specified by the customer,” Cram said.

Data centers can build their own PDU using the company’s website.

The company’s PDU solutions are supported by the major data center infrastructure management software tools, both proprietary and open source. That means, Cram said, that software developers can remotely reboot locked-up servers located in cloud data centers. IT managers can turn off unused outlets and require that new hardware be installed to request a ticket to enable power for the device.

“Data center managers use our products to help load balance across power phases and to determine their available capacity within the cabinet, row or the entire data center,” he said.

Renewable focus

Renewable energy is becoming a big focus for data centers. Wind, solar, hydroelectric, tidal, biogas-powered fuel cells and geothermal options are all recent possibilities. They generate DC or AC power.

“Getting DC power into the data center is different from AC power. Hence, many solar installations turn their DC-output into AC to put it onto the grid before transport to the data center consumers,” Cram said.

Legrand, Server Technology’s parent company, is expanding its portfolio of power-related solutions for the data center in both AC and DC power. In addition to its AC offering from 120-volt (V) single-phase up through 480V three-phase, Server Technology offers -48V DC and some 380V DC solutions.

“Contractors should always keep their eyes and ears open for opportunities to solve problems for their data center customers. Solving those problems may involve making use of the wide array of sensors available today,” Cram said.

Growing density and PDUs

In the meantime, rack-power density will continue to increase, driven by the very applications that help centers manage their own power—such as AI and machine learning, high-performance computing and big data. Today, the average power consumption, at any given time, for a rack is around 7 kilowatts (kW), depending on the data center. However, almost two-thirds of data centers in the United States experience higher peak demands, with a power density of around 15 or 16 kW per rack. Some data centers can reach 20 or more kilowatts per rack.

“As demand grows, the challenge will be to provide power where it is needed reliably and efficiently,” Mott said.

That’s why data centers are moving to intelligent rack PDUs and sensors, Mott said. PDUs can provide remote power monitoring of current, voltage, power and power factor and energy consumption.

“When power data is collected at the granular level, it can open up tremendous insights about phase imbalances, circuits becoming overloaded and more,” Mott said, which is where AI comes in. The data can be used to plan and prevent unnecessary downtime and identify areas of excess capacity.

Metering at the inlet, outlet and circuit levels can help determine the power usage across an entire data center.

“Inlet metering, for example, is crucial for determining overall server power usage and availability at the rack,” Mott said.

Additionally, metering at the outlet can help centers understand power consumption of a specific device or server, while metering at the circuit breaker provides early warnings if a circuit becomes heavily loaded and runs the risk of tripping.

“Raritan’s intelligent PDUs offer more than just power distribution, they are the starting point to gaining valuable insights about the environmental health of a data center,” Mott said.

Makes good sense

ECs should look for smart sensors that can be deployed as plug-and-play options that work with intelligent rack PDUs, inline meters, branch-circuit monitors or other solutions in order to eliminate the need, and cost, for a separate controller and underlying networking, Mott said.

Mott also recommends choosing sensors that have detachable, replaceable sensor heads that make it easy to maintain accuracy.

“If a sensor’s accuracy diminishes, you don’t need to remove the entire sensor; just replace the sensor head to maintain a high degree of accuracy.”

About The Author

SWEDBERG is a freelance writer based in western Washington. She can be reached at [email protected].





featured Video


New from Lutron: Lumaris tape light

Want an easier way to do tunable white tape light?


Related Articles