Skip to content

Storage Data Path#

conceptstub

What it is#

The path that feeds data and models to GPUs and absorbs checkpoints: object storage, parallel file systems (Lustre, Weka, GPFS), local NVMe caches, and GPUDirect Storage.

Why it matters#

GPUs that wait for data are wasted. Training reads large datasets continuously and writes multi-terabyte checkpoints in bursts; inference loads large model weights at startup. Storage runs in parallel with the network data path, so it is designed alongside the fabric, not after it.

In this module#

A data path parallel to L6–L9: feeding GPUs and absorbing checkpoints.

Planned content

  • Data path: object storage → parallel FS → local NVMe → GPU
  • CSI drivers: local NVMe, Rook/Ceph, Lustre, Weka
  • Checkpoint bursts and spot/preemptible recovery
  • GPUDirect Storage