AI Drives Meituans Performance and Growth
/w=1080,quality=90,fit=scale-down)
Background
Meituan stands at the forefront of AI innovation in China's e-commerce sector, leveraging artificial intelligence to revolutionize its lifestyle services platform. As its user base expands and business demands evolve, Meituan tackles complex AI challenges head-on, pushing the boundaries of system performance. The company recognizes AI as a critical driver of business innovation in the fiercely competitive digital marketplace. By harnessing advanced AI technologies, Meituan enhances the accuracy and personalization of its internet services, solidifying its position as a leader in AI-driven e-commerce solutions. This strategic focus on AI empowers Meituan to meet the sophisticated demands of modern consumers while continuously advancing its technological capabilities.
Executive Summary
- Meituan integrated AI across its operations, from site selection to marketing
- Rapid growth in data and AI model complexity necessitated infrastructure re-architecture
- Partnership with Intel leveraged Intel® Xeon® Scalable processors to optimize TensorFlow applications
- Achieved 10x improvement in distributed scalability for recommendation systems
- Significant performance gains realized in TensorFlow model training
- Challenges included TensorFlow performance bottlenecks and rising total cost of ownership
Problem
Meituan's rapid growth and increasing reliance on AI-driven solutions led to significant challenges. As the scale and complexity of their AI models grew, performance bottlenecks in TensorFlow became apparent, particularly in large-scale applications. The official version of TensorFlow exhibited issues such as memory resource wastage and scalability concerns. Additionally, the rising total cost of ownership (TCO) posed a significant challenge, necessitating a balance between performance improvements and cost considerations.
Solution
To address these challenges, Meituan embarked on a comprehensive re-architecture of its infrastructure and optimization of its software. The company partnered with Intel, integrating Intel® Xeon® Scalable processors into its TensorFlow system. These processors, known for their AI acceleration capabilities, became a crucial component of Meituan's server clusters. The optimization process involved multiple stages, from enhancing throughput for unit computing power to improving distributed computing. Techniques employed included utilizing Intel's Advanced Vector Extensions and parallelization technology, optimizing load balancing, and enhancing communication mechanisms.
Impact
The optimizations resulted in substantial improvements in system performance. Meituan achieved near-linear acceleration for models with hundreds of billions of parameters, enabling a full year of sample data training in just one day. The TensorFlow-based Recommendation System became more cost-effective and efficient. Furthermore, the enhanced user-friendliness of the system led to its adoption across multiple Meituan businesses, including food delivery, ad platforms, and community group buying.
Change Management
Implementing these changes required a coordinated effort across Meituan's organization. The company built a robust AI team dedicated to infusing AI-driven solutions into its operations. This team worked closely with various business units to ensure smooth integration of the optimized TensorFlow system. The transition involved training staff on the new infrastructure and optimizing workflows to leverage the improved performance capabilities.
Roadmap
Looking ahead, Meituan remains committed to its mission of helping people lead better lives through continuous investments in R&D and digital transformation. The company is poised for even greater successes in the AI-driven digital age. Meanwhile, Intel's ongoing commitment to advancing AI technology, in collaboration with partners like Meituan, aims to drive further development in the AI industry. The focus is on delivering more accurate, personalized services to end-users while making AI technology more accessible and cost-effective.
Publication: