Packing dataset files into sequential blobs lifted one training pipeline from 36 to 47 steps per minute

irinarish · x · 2026-07-21

- The author sped up an ML training pipeline from **36 steps/min to 47 steps/min** by changing how datasets are stored on disk. - On consumer hardware, storage is often the bottleneck: **NVMe is expensive**, while **SATA HDDs are slow**, especially for datasets made of many small files. - The fix was to **pack files into sequential compressed blobs** such as **TAR, Parquet, or WebDataset** so the drive can do mostly sequential reads instead of random seeks. - If the data is already compressed, storing it **uncompressed** can avoid extra compression/decompression overhead; the author cites **FLAC** as an example. - The post also notes that very large training runs often rely on **network storage blocks in the same datacenter** as the GPUs, because the dataset is too large to keep on device.

Original post →

More from Infra

Infra channel →