Background

Migu is China Mobile’s specialized subsidiary for the mobile internet, responsible for product provision, operations, and services in digital content. It is the sole operating entity for China Mobile’s music, video, reading, games, and animation digital businesses, with five subsidiaries: Migu Music, Migu Video, Migu Digital Media, Migu Interactive Entertainment, and Migu Animation.

Migu has become a leading all-scenario brand immersion platform in China, gathering more than 17 million songs, 4.30 million videos, 1,200+ audio and video live channels, 500,000+ books and periodicals, 30,000+ games, and 470,000 episodes of animation. On this vast volume of data, Migu carries most of China Mobile’s artificial-intelligence business. To better carry out AI R&D and application, Migu procured GPU servers and related software to build an AI-domain technology platform.

Requirements

Migu’s centralized procurement of rack GPU servers is to meet large-scale data-computing requirements of Migu’s IT resource-pool project. The main construction is an ultra-large-scale offline AI training platform, an online inference platform, and a corresponding cloud-computing platform, with a focus on R&D of face recognition, intelligent customer service, public-opinion analysis, and other AI technologies and applications suited to the communications industry. GPU-server suppliers make targeted omissions and optimizations against GPU-server configuration and management requirements.

Rack GPU servers should, in line with the requirements of large cloud data centers today, simplify on-site maintenance and raise support for remote, automated operations management.

Solution

GPU servers in Sugon’s AI product series are a class of GPU servers aimed at medium-to-high power-density data centers and standard 19-inch racks, with flexible procurement and deployment.

Typical configurations 1 and 2 used 4U 8-GPU servers with four V100 and four P40 GPUs respectively, plus dual-port 25GE fiber NICs supporting RoCE, raising device information-processing bandwidth and lowering transmission latency, mainly for deep-learning scenarios in artificial intelligence.

System-stability tests and GPU-card performance tests were conducted on GPU cards that met the requirements together with the GPU servers chosen for this project, and related test methods and reports were provided, strongly verifying product stability and high performance.

Typical configurations 3 and 4 used 4U 4-GPU servers with four P40 and four P4 GPUs respectively, mainly for online inference and video encoding/decoding scenarios in artificial intelligence.

Sugon developed a deep understanding of Migu’s AI applications, helped build AI training models, continually proposed optimizations, and shared Sugon SothisAI technology. Sugon SothisAI is a cloud platform dedicated to deep learning. It embeds deep-learning frameworks such as Caffe and TensorFlow, integrates a task-scheduling system, and, combined with Docker container technology, provides users with deep-learning compute services, concentrating dataset management, image management, container management, model management, file management, task management, and resource management. It schedules and allocates high-performance computing resources, submits training tasks, manages tasks, and monitors resource status, providing an integrated solution for deep-learning clusters.

It was also paired with Sugon cluster-management software, giving Migu’s AI-domain construction all-round industrial design, job scheduling, and cluster monitoring and management, with convenient application-software services that make industrial design simpler, powerful job scheduling that makes computing more efficient, and rich cluster configuration and management tools that simplify cluster management. Fine-grained presentation of cluster running status and timely alerts on abnormalities help prevent hidden issues. The system presents the running status of all types of software and hardware resources intuitively, locates device fault sources accurately and quickly, and keeps IT equipment running safely and stably.