Storage device specification database#

The benchmark directory contains two normalized JSON tables:

The specifications come from manufacturer product pages or data sheets. Unknown values are null; reseller claims are not used to fill gaps. A raw label that does not uniquely identify physical hardware also has a null spec_id. This is important for QEMU, generic nvme, and OEM labels whose physical backing device is not visible.

The files are intentionally flat arrays so DuckDB can join them directly:

WITH aliases AS (
  SELECT *
  FROM read_json_auto('orchestrator/benchmarks/storage-device-aliases.json')
),
specs AS (
  SELECT *
  FROM read_json_auto('orchestrator/benchmarks/storage-device-specs.json')
)
SELECT
  run.*,
  aliases.spec_id,
  specs.product_family,
  specs.interface,
  specs.rated_sequential_read_mb_s,
  specs.rated_random_read_iops,
  specs.nand,
  specs.host_memory_buffer
FROM run
LEFT JOIN aliases ON run.advertised_disk_name = aliases.raw_label
LEFT JOIN specs USING (spec_id);

The executable RTX 5090 official-spec analysis applies the same join to the dashboard's latest-host API. It prefers an exact advertised label, then falls back to the Linux NVMe model only when exactly one NVMe is visible; it does not guess which device backs a multi-device container. The query reports rated-to-realized and realized-to-application Pearson and average-tie Spearman correlations plus the per-host realization ratio.

Vendor ratings are explanatory covariates, not observed benchmark results. They are typically fresh-drive, high-queue, best-case maxima and do not describe Vast's filesystem, container backing path, page-cache state, block queue, PCIe attachment, thermals or contention. The measured model reader and in-workload block counters remain the primary evidence.

When a new raw model appears, add an alias row first. Add a specification row only when an official source identifies the exact family and capacity. Record the official URL and retrieval date, and leave unpublished fields null.