How to solve the problem?
In order to meet all the challenges, China Unicom has chosen the Chronos framework provided by Intel to complete the above process quickly and efficiently. The framework, derived from BigDL (Intel’s open-source unified Big Data analytics and AI platform), provides users with three capabilities:
▪ Data Processing & Feature Engineering components: It has more than 70 built-in data processing and feature engineering tools, which can be easily called by the TSDataset API interface, so as to complete the data preprocessing and feature engineering process quickly and efficiently.
▪ Built in Models: More than 10 independent deep learning and machine learning models from Forecasters, Detectors and Simulators for time series prediction, detection and simulation;
▪ Optional HPO components: Help China Unicom complete the automated machine learning process with a highly integrated, extensible and automated workflow (through apis such as AutoTSEstimator). Intel optimizations, such as the integrated ONNX runtime and Intel ® oneAPI AI Analytics Toolkit, provide support for reasoning.
The second step of using network AI to achieve energy saving and emission reduction is to construct an effective 5GC network element resource occupancy prediction scheme
Based on Chronos framework, it becomes convenient and methodical for China Unicom to construct 5GC network element resource occupancy prediction scheme. The whole process can be divided into the following steps:
▪ Used historical business data (such as traffic data, etc.) and server resource utilization logs (such as processor occupancy rate, etc.) for modeling, and TSDataset API interface provided by Chronos framework quickly implemented filling, scaling and other operations on the time series data, and carried out automatic feature generation;
▪ Conducted hyperparameter search through API such as AutoTSEstimator and retrieved the optimal hyperparameter set according to the prediction target, then optimized the model and data processing process and formed the time series prediction model;
▪ Use this model to reason (or evaluate and optimize) real-time business data to obtain final processor occupancy projections.
At present, the new scheme has been tested in the 5GC test resource pool of China Unicom, and the final MSE result of the comparison between the predicted result of processor occupancy and the actual value is only 1.71, while the actual absolute error of the predicted processor occupancy is less than 1.4% on average, which meets the expectation of China Unicom for the new scheme.
So what are the benefits of using this prediction scheme for China Unicom’s green energy saving? According to a calculation made by China Unicom’s experts, the predicted solution combined with the processor frequency reduction technology is expected to reduce the energy consumption of a single server by more than 15%, which translates to the overall cloud resource pool and can directly save 46 million KWH of energy every year. Combined with other energy-saving measures, it can reduce carbon dioxide emissions by about 60,000 tons per year.
It can be seen that the construction of green data centers in the future may not only reduce the “mountain path” of PUE, but also create a new way to save energy and reduce emissions in the ICT field with the help of AI solutions brought by the trend of network intelligence. Intel is also making continuous efforts in this direction, providing comprehensive support from computing facilities to AI framework for the vast number of telecom operators, equipment manufacturers and cloud service providers, including China Unicom, to help achieve a larger range and scale of energy conservation and emission reduction effects, and to implement the construction of “new information infrastructure for green data centers”.
Post time: 11-25-2022