Efficient Performance-Aware GPU Sharing with Compatibility and Isolation through Kernel Space Interception

Learn about KRYPTON, a novel kernel-space GPU-sharing scheme that addresses critical limitations in existing GPU virtualization approaches through this 14-minute conference presentation from USENIX ATC '25. Discover how current GPU-sharing solutions suffer from compatibility restrictions limited to CUDA runtimes and vulnerable isolation mechanisms where single application failures can disrupt all other applications sharing the same GPU. Explore the research findings that reveal these limitations stem from user-space designs of existing solutions and understand how KRYPTON's kernel-level interception of GPU command buffers provides virtual GPU devices with robust compatibility and isolation. Examine the innovative spatio-temporal sharing approach that replaces fixed GPU resource allocation to enable performance guarantees while maximizing resource utilization. Analyze experimental results demonstrating KRYPTON's ability to reduce required GPU numbers by 32.1% compared to state-of-the-art baselines while maintaining strong compatibility and isolation properties, presented by researchers from Shanghai Jiao Tong University and Lenovo.