How artificial intelligence makes datacenters more efficient

Can artificial intelligence play a role in an optimal design for modern datacenters? Martin Matse from Perf-iT thinks so. He states that AI can be an important tool for implementing Data Center Infrastructure Management (DCIM). During the IT Infra event, Matse delves deeper into the benefits of applying artificial intelligence.

November 6, 2019 By: Dimitri Reijerman (translated from original Dutch text to English by Metho de V)

At a time where data center energy consumption is increasing rapidly, optimization is a must. But it is not self-evident that, apparent obvious technical adjustments actually benefits the datacenter. Matse: “Datacenters must continue to optimize. The magic value here is the PUE (Power Usage Effectiveness). But how exactly are you going to do that? And what is the result of your optimization? For example, you can lower the cooling water temperature by two degrees, because then you might save money. But what is the impact of this on the data floor? That is hard for a person to quanitfy. “

Artificial intelligence, a technology that is developing rapidly and has more and more applications, can therefore be a useful tool to improve the availability and efficiency of a datacenter. Matse says: “What we do, together with TNO (Dutch Governmental Sciunce institute), Vortech (data science) and Actiflow (CFD), is making a ‘digital twin’. This is a digital copy of reality that also incorporates physical algorithms of a datacenter, such as air flow and temperatures. The great thing about such digital twins is that you can test different scenarios based on data. Based on the new data, the AI model will tell exactly what the impact will be on the data floor.”

Sensors & algorithms

Preparing a working model requires the necessary preparation, Matse says: “What we do now: we make a basic model and mount sensors in various places. The model has a certain expectation based on the input. And based on the sensors, this model sees what the actual situation is on the data floor all the time. The algorithms become more and more accurate this way: the model continuously adapts to the newly found reality.”

The use of artificial intelligence for DCIM is still relatively new. That is why many tests are still being done with it, Matse says: “TNO has developed and patented the basic technology. We have now tested it in practice at two different data centers: Infrabel (the Belgian railway infrastructure provider) and at NL-DC. Applying this new DCIM technology results in a very fast ROI. Things happen on the data floor that were not expected beforehand. “

To get the model better and better, many operating hours are needed: “A model is dynamic,” says Matse. “Based on the process and the placement of more and more sensors, you can see that the model is becoming more accurate over time. The best-known forms of AI, deep learning and machine learning, cannot make predictions outside of the data it has ever seen. But a digital twin goes one step further, because you have also recorded physical processes in its algorithms through real-life sensor data. “

At present there are few companies that offer DCIM on the market based on machine learning and deep learning or apply it in their computer centers. According to Matse, this is due to the aforementioned limitations of these implementations of artificial intelligence. He gives an example: “In the US, they used deep learning to select applicants. They have conducted the AI with countless CVs of ideal candidates. All women were then rejected. How did that happen? There were no women’s CVs in the input data. So, the completeness of your data is therefore important. This is less important with our model because you actually use algorithms as the basis of your model.”

Future of DCIM and AI

In the future, applying AI for DCIM applications will still encounter the necessary challenges: “The challenge is primarily acceptance. It is still very scary for many people to have a computer model determine how a data center can be optimally controlled. But it is becoming more common. We will often first see an intermediate step: that the model provides advice and does not actively manage the data center itself.”

Finally, Matse also points to the social advantages of this technology, in particular in the area of sustainability: “I think that there is still a lot to be achieved within data centers to reduce energy consumption. That is an issue throughout the country and especially in Amsterdam and its surrounding area. The energy transition is a nice by-catch at DCIM, for example for the reuse of heat from data centers. If you cool cleverly, the return air temperature also goes up. This increases the efficiency of your heat pumps. When municipalities become very strict to enforce the use of residual heat, like for example at the construction of a new district in Amsterdam West, you can hardly ignore using ‘waste heat’ coming from datacenters.”

Follow us, or subscribe to our newsletter ( 3 to 4 x per year)