Problem Description

When running containers with the --gpus all option, optimized images created by the optimize command show poor prefetch performance with significant prefetch misses.

Root Cause Analysis

The issue stems from a timing gap in file access tracking:

1. Pre-container OCI Hook Execution: NVIDIA runtime injects dynamic libraries and executes ldconfig via OCI hooks before the container actually starts
2. Early File Access: The ldconfig operation accesses various dynamic libraries in the image during this pre-container phase
3. Missing Tracking: The current optimize process only starts fanotify monitoring after the container is running, missing the file access patterns from steps 1-2
4. Runtime Prefetch Misses: When the container runs, the same OCI hook behavior repeats, causing FUSE to trigger on-demand loading for files that should have been prefetched

Proposed Solution

Implement a pre-container monitor that:

• Starts tracking file access before container creation
• Use FAN_MARK_FILESYSTEM to run fanotify on the host. (Testing has shown that FAN_MARK_MOUNT is unable to track a container's rootfs).
• Captures access patterns from OCI hooks and pre-container operations