DGraphFin Force-SSD Read-Amplification Report

Real device I/O measured per page layout. Cross-check these numbers with your own eBPF/blktrace/iostat counters.

dataset synthetic-smoke · O_DIRECT · layouts natural,random · fanout 10,5 · batch 128 · pages/batch 8 · replicas 1 · generated 2026-06-02 13:32

The value: locality lowers device reads

The neighbor pattern expands random training seeds to their sampled neighborhood. A locality-aware layout (metis) keeps those neighbors on a few 4 KiB pages, so the device serves far fewer reads per useful byte. The measured O_DIRECT amplification tracks the in-RAM modeled metric, which is exactly the agreement an external observer should confirm.

The gap: the supervised-read floor

The page-aligned sweep reads whole training pages. RA_fetch is ~1x for any layout, but only a fraction of nodes per page are supervised, so RA_signal — bytes read per supervised node — cannot drop below the ~4.3x floor no matter how good the layout is. That floor is the gap the layout work leaves open.

All metrics

Verify it yourself

We do not run eBPF here; the harness only issues the reads. Attach biosnoop/biolatency, blktrace or iostat -x to the device backing the store and confirm the per-I/O read intent and amplification independently.