etcd on the Intel DC S3610 — October 11, 2026¶
What this is: measurements from moving etcd on the HP SFF control plane to an
enterprise SSD. Use them as the baseline when new control planes join. The current layout is in the
disk map; this page keeps the numbers behind it.
Raw data: docs/assets/fleet-benchmarks/2026-10-11/.
etcd on the Intel DC S3610¶
The control-plane VM disk moved from a consumer PNY CS900 to an Intel DC S3610
400 GB (enterprise SATA, power-loss protection) with an online pvmove.
Same VM, same etcd member, same workload.
etcd's own latency histograms (Prometheus, 6-hour windows, Oct 1–11):
| Metric | PNY CS900 | Intel S3610 | Change |
|---|---|---|---|
| WAL fsync p50 | 1.66 ms | 0.56 ms | 3× faster |
| WAL fsync p99 | about 7.0 ms | 3.8–4.4 ms | about 1.8× faster |
| Backend commit p50 | 3.05 ms | 0.61 ms | 5× faster |
| Backend commit p99 | about 7.6 ms | 3.8–5.1 ms | about 1.7× faster |
| Mean physical write wait (host, 30 s sample) | 2.02 ms | 0.08 ms | — |
etcd's guidance is WAL p99 below 10 ms and backend commit p99 below 25 ms. Both were already met; the S3610 adds headroom for load spikes. The p99 rose slightly over the last day. etcd's database file is 412 MB with about 15% live data, so defragment it in a quiet window before blaming the disk.