SK Telecom: AI Pipeline Improves Network Quality

Accelerated end-to-end network AI pipelines use Analytics Zoo, TensorFlow, and Apache Spark on Intel® architecture.

At a glance:

  • SK Telecom is the largest mobile operator in Korea. Building on its strength in mobile services, the company is also creating value in media, security, and commerce.

  • To effectively analyze the massive amount of data their network generates, SK Telecom and Intel engineers built an end-to-end network AI pipeline for network quality prediction. The entire pipeline runs on a unified Intel® Xeon® Scalable processor-based server cluster with Intel® Advanced Vector Extensions 512 and Intel® Deep Learning Boost, while Analytics Zoo software handles the in-memory data pipelines and distributed model training and inferencing.

author-image

โดย

Executive Summary

Network quality is becoming harder for communications service providers (CoSPs) to manage manually because of the overwhelming flood of telemetry data coming from multi-gigabit networks. This challenge grows with the rapid advancement of 5G technology due to the large number of devices and very fast data rates. As a result, managing communication networks in an intelligent and automated fashion using artificial intelligence (AI) technology becomes increasingly important for CoSPs.

SK Telecom, the largest mobile operator in South Korea, manages more than 400,000 cell towers with over 27 million subscribers. This network handles 1.4 million records every second, which accumulates to 120 billion records per day.1 In order to effectively analyze this massive amount of data, SK Telecom and Intel engineers built an end-to-end network AI pipeline for network quality prediction using Analytics Zoo and FlashBase, running on Intel® architecture servers, which effectively applies a memory-augmented TensorFlow model to large-scale time series data on Apache Spark.

The entire pipeline (from FlashBase to Spark DataFrames to TensorFlow) runs on a unified Intel® Xeon® Scalable processor-based server cluster, with Intel® Advanced Vector Extensions 512 (Intel® AVX-512) and Intel® Deep Learning Boost. Additionally, this leverages Analytics Zoo software to automatically handle the in-memory data pipelines and distributed model training and inferencing. In tests conducted by SKT, this AI pipeline outperforms SKT’s legacy GPU-based implementation by up to four times and six times for deep learning training and inference respectively,2 which enables SK Telecom to more quickly forecast and detect degradation and abnormal changes in network quality so that SKT can take proactive action to deliver their 5G service quality.

Read the white paper – SK Telecom, Intel Build AI Pipeline to Improve Network Quality

ข้อมูลผลิตภัณฑ์และประสิทธิภาพ

1Data from SK Telecom, September 2020.
2Tests conducted by SK Telecom in Feb. 2020: The Analytics Zoo server was an Intel® Server System R2208WFTZSR powered by a 2.6 GHz Intel Xeon Gold 6240 processor (microcode 0x400002c). The server featured three nodes and six sockets. Both Intel® Hyper-Threading Technology and Intel® Turbo Boost Technology were turned on. Total memory equaled 256 GB. CentOS 7.8 (kernel 3.10.0) was the operating system and the server ran the SK Telecom Lightning DB application. Other software included Analytics Zoo v0.7, Tensorflow v1.15, Pandas v0.25.3, NumPy v1.18.0, and Dask v2.7.0. The GPU server was a HPE DL380 Gen 9 powered by a 2.4 GHz Intel Xeon E5-2680 v4 processor (microcode 0xb00001e) and an Nvidia P100 GPU (AI training)/K80 (AI inference). The server featured one node and two sockets. Both Intel Hyper-Threading Technology and Intel Turbo Boost Technology were turned on. Total memory equaled 256 GB. CentOS 7.3 (kernel 3.10.0) was the operating system and the server ran the SK Telecom Lightning DB application. Other software included Tensorflow GPU v1.12, Pandas v0.25.1, NumPy v1.14.5, and Dask v2.7.0.