container performance

Containers can quietly sap 10-40% of AI speed; this guide shows how to reclaim near-bare-metal performance on Linux with lean, GPU-aware images, correct NVIDIA passthrough, tuned CPU/memory/NUMA/IPC (cpusets, /dev/shm, HugePages, threads), faster I/O and networking via NVMe bind-mounts and host networking, plus a measure-iterate loop, with Docker/Podman commands, pitfalls, and a Triton inference quick start.