Diagnosing storage bottlenecks
System-level I/O diagnostics:
iostat -x 1 # Per-device I/O stats, 1-sec interval
iotop # Per-process I/O usage
dstat # Combined system stats
# Lustre filesystem
lfs df # Filesystem usage
lfs df -h # Human-readable
lfs setstripe -c 4 /path/to/file # Set stripe count across 4 OSTs
Signs of a storage bottleneck:
- High I/O wait time (
%iowaitiniostatortop) - GPU utilization drops while data is loading (GPU starved)
- Training throughput limited by data pipeline, not GPU compute
