As of September 28, 2022, the ICT (Information and Communication technology) industry’s share of global carbon emissions is no longer negligible. With the proposed carbon neutrality goal of 2060 in China, the majority of telecom operators, equipment providers and cloud service providers, including China Unicom, have also taken the construction of green energy-saving data centers as the focus of future network infrastructure construction. PUE (Power Usage Effectiveness, Energy efficiency) has attracted a lot of attention.
To promote the energy consumption management of computing power center requires a change of thinking
PUE is the ratio between the total energy consumption of a data center and the energy consumption of IT equipment. Currently, the PUE value of mainstream data centers generally reaches 1.x, where 1 can be regarded as the energy consumption of IT equipment, and X is the additional energy consumption caused by cooling and power supply of IT equipment. As the PUE value is the key index to measure the green energy saving level of data center, it has long been the goal of data center construction and upgrading to reduce the PUE value by using the limit of heat dissipation efficiency of various schemes. However, the pressure drop space of the PUE value is obviously shrinking, and the closer it gets to the limit value 1, the more cost pressure it will face in construction and maintenance, and even greatly exceed the reasonable budget, resulting in an embarrassing situation of saving energy but not saving money.
From the perspective of China Unicom, the ultimate goal of green data center is to reduce the overall energy consumption. The pursuit of the reduction of X is certainly one direction, but when the reduction of X meets the bottleneck, can we start with the energy consumption of IT equipment and make it feasible to change 1 into 0.9 and 0.8? In the view of China Unicom experts, although technically feasible, it is very risky. The reduction of energy consumption of IT equipment will inevitably lead to a decline in processing performance, which is likely to affect the service quality and even availability of the business for the 5G communication network that provides real-time voice and data services to hundreds of millions of users, resulting in unpredictable losses.
Then, can we control energy consumption on the premise of ensuring that the computing demand and service quality meet the standards by perceiving the computing demand and service quality index of upper business in real time? China Unicom and Intel will cooperate in this regard, combining China Unicom’s rich experience in network operation and Intel’s comprehensive and rich AI+ energy saving technology, to jointly create business-oriented intelligent energy saving solutions.
Network intelligence + Intel product technology combination, to predict the model to achieve new solutions of energy saving and emission reduction
The most effective way to understand the computing needs of a business in real time is to build a predictive model. Nowadays, the network intelligent technology, which has gradually become the trend of the industry, makes the model building step forward a solid step.
The so-called intelligent network technology, in short, is the integration of network and AI innovation, it is through the deep integration of AI and communication network hardware, software, system, etc., to present more intelligent features for network operation and service. One important point is that users can realize more effective perception, collection, processing and feedback of all kinds of network data through AI capabilities. Taking advantage of this feature, China Unicom is using data on various business loads in its data centres to create new forecasting models.
For example, the service load such as traffic is a typical time series model, that is, the service load changes in real time and has a time sequence. In contrast, the business load demands server resources such as processors and memory. Therefore, to build a forecast scheme, as long as it is clear which server resources are closely related to the service volume, the dynamic resource adjustment scheme can be developed according to the forecast results without affecting the continuity of the service.
Intel hardware products have a series of advanced features, for the server energy saving provides flexible adjustment capabilities. The Intel-based hardware infrastructure deployed in China Unicom’s data centers provides a built-in power mediation mechanism to fine-tune Core and Uncore frequencies according to different service loads. Power consumption can be further reduced on the premise that SLA levels are met.
Meanwhile, in the opinion of Xia Lei, chief engineer and chief architect of artificial intelligence at Intel, thanks to the full integration of AI technology, advanced hardware and software equipment and data, various kinds of real-time analysis and prediction AI applications arising from the rise of network intelligence will naturally become the help for performance optimization and active energy consumption management of 5G networks.
Now China Unicom is working with Intel to create a new energy consumption optimization scheme for data centers by utilizing the Chronos framework built based on BigDL components, accurately forecasting and refined management of resource demands, and “counting every bit” of server energy consumption in a dynamic adjustment way, so as to effectively reduce the overall energy consumption of data centers.
The first step to using network AI to save energy and reduce emissions is to choose a more efficient AI framework
Traditionally, it is not easy to build a time series prediction model and form an efficient and usable AI application, because it involves a series of steps from data acquisition and preprocessing, feature engineering and model training. If China Unicom engineers had to design and build each step from scratch, it would be time-consuming and laborious.
At the same time, in order to improve the accuracy and performance of the model, the construction process often needs to spend huge manpower and time resources to manually optimize the Hyperparameter to achieve more efficient Hyperparameter optimization (HPO), which is also an important challenge facing China Unicom.
Post time: 11-17-2022