GPU 클러스터를 아껴쓰고 나눠쓰고 바꿔쓰고 다시쓰기 - 동적 할당 기술과 고성능 클러스터링 구현

Explore advanced GPU cluster optimization techniques in this 23-minute conference talk from SK AI SUMMIT 2024. Learn how to implement dynamic GPU allocation technology that enables multiple users to instantly share a single large GPU cluster according to their needs. Discover solutions to performance degradation issues commonly found in virtualized environments and understand the technical challenges and solutions involved in implementing high-performance GPU clustering at physical equipment levels. Examine real-world case studies from Japanese cloud companies and domestic public AI projects that demonstrate operational efficiency improvements and cost reduction effects. Gain insights into network infrastructure research and development for cloud environments, including analysis and resolution of network bottlenecks and failures in complex large-scale cloud systems. Master optimization techniques across the entire stack from hardware offloading to kernel, container, and VM networks to ensure stable operation of services requiring extreme performance like GPU clusters.