how to solve the problem of lack of electricity for IT infrastructure
The other day we discussed the work of Internet providers, and also talked about problems with cloud regulation. Let's continue the last topic, but from the point of view of IT infrastructure. The fact is that the world is experiencing some difficulties with the construction of data centers, and the energy infrastructure cannot cope with the load – blackouts in data centers are becoming more and more destructive (although more rare).
![Photo: Severin Höin / Unsplash Photo: Severin Höin / Unsplash](https://habrastorage.org/getpro/habr/upload_files/1f7/337/a2c/1f7337a2ce36dccb68542bb72c3c6785.jpeg)
Growing Appetite
Companies are building increasingly large-scale data centers. For example, in the US state of Arizona there is a data center with an area of 35 thousand square meters. m. However, building a huge machine room is only half the battle; then it needs to be “fed.” Data center mentioned accommodates thousands of computers with a planned capacity of over 280+ MW, which is equivalent to the energy costs of thousands of homes. And there are not many sites left in the world suitable for launching data centers with similar appetites, that is, sites with affordable and sufficiently powerful energy supplies.
By data International Energy Agency (IEA), the world's data centers were already consuming 1% of the world's electricity in 2021, and that figure has definitely increased since then. Operators today face energy shortages. British electricity company Aggreko interviewed more than 700 specialists from data center operating companies. According to the report, many of them not surethat they will be able to provide their facilities with electricity and maintain them in working order even in the next five years.
One striking example is the situation with AWS in Europe. Due to growing demand for cloud services and energy constraints, the company was forced to introduce limits for the deployment of GPU nodes in an Irish data center. This decision most strongly affected high-performance computing and tasks related to training ML models. High load on the power grid in general has meant that Irish data centers are having to redistribute the load to data centers in Sweden and other parts of the EU.
Why is this happening
The most obvious reason for the increased workload is the digitalization of almost all areas of life. It has led to an exponential increase in the volume of data that needs to be stored somewhere and an increase in the load on the infrastructure of data centers. According to Statista forecasts, the volume of data generated in 2024 will increase by 22.5% compared to the previous year and will amount to 140 zettabytes.
The second reason is to gain access to reliable sources of electricity. it gets more difficult. There are fewer and fewer available sites for data centers near large power plants. Prices for land plots that meet the demands of operating companies are soaring, which also creates additional problems during the construction of data centers. According to data provided by the European company Aggreko, demand in the data center market is outpacing supply. But arise difficulties caused by rising prices for materials and a lack of qualified personnel. As a result, the construction of facilities is delayed, and startup errors occur, which undermine the stability of data centers.
![Photo: Fré Sonneveld/Unsplash Photo: Fré Sonneveld/Unsplash](https://habrastorage.org/getpro/habr/upload_files/765/62e/fbd/76562efbd967ff1d6c6808e0b6cb6497.jpeg)
The development of artificial intelligence systems is further exacerbating the power shortage problems of data centers. Training modern ML models requires enormous computing power. CPU and GPU in the LLM learning process can consume more than a thousand megawatt hours. Engineer Alex de Vries from the Free University of Amsterdam suggests that in the near future, humanity will begin to spend about 30 terawatt-hours annually on maintaining the operation of AI systems (this number is comparable to the energy consumption of Ireland). Colleagues from the University of Florence believe that the forecast may even be understated.
How to solve the problem
Cloud giants choose a simple, but not the cheapest method – they repurchase complexes with sufficient power for their needs. So, at the beginning of the year, one large cloud provider bought it out a complex of data centers worth $650 million. Interestingly, it was built next to the 2.5 gigawatt Susquehanna nuclear power plant, located in northeastern Pennsylvania.
At the same time, to solve the problem of energy efficiency, data center operators use demand management mechanism for electricity – demand-side response (DSR). This is a scheme in which data centers modify the level of energy consumption, depending on the capabilities of the local power plant. Thus, if the load on the network increases sharply, the data centers connected to it partially switch to power from generators and spare batteries. At the same time, data center operators can sell the excess electricity they have accumulated into the network. For example, DSR is active uses Microsoft at its data center outside Dublin.
Looking to the future, the market is seeing growing interest to a new form of nuclear energy known as small modular reactors (SMRs). SMRs are significantly more compact and produce less energy than traditional reactors. However, they are also cheaper compared to other nuclear energy projects. Unfortunately, modular reactors are still only promising solution, which is in the early stages of implementation. According to analysts, data centers will come to this format of energy supply in 10–15 years.
There is also an opinion that compact and open machine learning models can become a solution to the problem of high energy consumption in data centers. They will reduce the amount of calculations, and at the same time reduce the load on data centers. Customizing language models for specific tasks also will allow reduce the cost of their service.
The growing load on data centers also requires improved equipment to increase energy efficiency. And Arm chips are already find application in data centers, despite the fact that previously the company’s technology was used to a greater extent for smartphones. In theory, they can reduce the energy consumption of computer rooms by more than 15%.
The situation with blackouts
Despite problems with electricity and lack of space for data center construction, statistics show that the number of blackouts in data centers is decreasing. But the consequences of such failures are becoming increasingly devastating. According to Uptime Institute, for each serious failure, data center operators lose from 100 thousand to 1 million dollars. We must not forget about reputational damage, which is not always possible to assess.
![Photo: israel palacio/Unsplash Photo: israel palacio/Unsplash](https://habrastorage.org/getpro/habr/upload_files/39f/15e/cf6/39f15ecf6182c058875893ec699c0dfa.jpeg)
For example, last year a large Western data center operator collided with a massive disruption to its cloud services due to a power outage. The blackout crippled infrastructure in Western Europe, causing companies using virtual machines and SQL databases to experience outages. An attempt to switch power to the generators was also unsuccessful, as some of them did not start.
In addition to generators, many data centers use lithium-ion batteries as backup power sources. Frost & Sullivan estimates that the share of such batteries in data centers was 15% in 2020, and is expected to grow to 38.5% by 2025. However, lithium-ion batteries are more dangerous in terms of fire compared to lead-acid batteries. When they break, flammable gases can be formed, which makes extinguishing fires difficult. The proliferation of these types of batteries may cause problems in the future if precautions are not taken.
Overall, blackouts are becoming an increasingly pressing issue for operators, requiring new solutions for reliable power supply to avoid major disruptions to critical infrastructure in the future.
What else to read on the topic and more:
Big Data Processing in QoE Stor. Our compact material, where we talk about the components of the solution, processing and storing information on the statistics server of our product for Internet providers.
Dynamics of DDoS attacks: overview of the first quarter of 2024. This is a continuation of our review material, in which we monitor how the geography and intensity of DDoS attacks are changing. We discuss the most notable cases – the spread of rapid reset attacks, as well as an attack on the infrastructure of the developer of open source tools for 3D modeling.
Forewarned is forearmed: a selection of open resources with information about identified vulnerabilities. We have prepared for you a compact list of databases where you can find up-to-date information about network vulnerabilities. Among them are the well-known CVE and NVD, as well as OpenCVE and VulDB. Each of them offers an API for sending and receiving notifications about threats – we'll tell you where to get the documentation.
Telecom versus streaming: how Internet providers are trying to attract investment in network infrastructure. Network equipment is becoming more expensive, and government support programs and subsidies are coming to an end. Providers are inventing new approaches to financing their activities. This is a traffic tax for large content platforms, mergers with partners and competitors, as well as new tariffs for subscriptions to TV series.
Build a topology and test fault tolerance: a selection of open source solutions for network emulation. Another selection with tools for training and practice of specialists in the field of network technologies. We talk about the capabilities of Containerlab, Mininet, Kathará and IMUNES and several other systems: what they can do, who developed them, what the community as a whole thinks about them.