Skip to content

perf.md

Latest commit

 

History

History
108 lines (93 loc) · 7.63 KB

perf.md

File metadata and controls

108 lines (93 loc) · 7.63 KB
Raw

Performance results

Status: complete curves; gates 12--13 and 16 fail, gate 14 passes

The final golden-image, default-4 MiB-chunk HTTP sweep completed against pgs3 and same-host MinIO with zero request or integrity errors. Gates 14 and 17 pass; fixed small-object gates 12--13 fail. Current golden-image SQL scale passes both LIST thresholds but fork takes 1944.071 ms, so gate 16 fails.

Evidence: final HTTP manifest, summary, environment, and raw samples; golden-image scale; and same-final-source, pre-golden-image scale iteration.

Environment and tuning

Field Recorded value
UTC interval 2026-08-31 06:56:08--06:56:51
Workspace unborn/dirty; SHA-256 780ea7bb4c2a852c60bec9b22e637ec051ddd43ed0fa21015d0107b166fd09b9
Host / Docker Apple M5 Max, 18 CPUs/128 GiB; Docker Desktop 29.4.1, 16 vCPUs/about 64 GiB, arm64
Baseline differences power-safe NVMe and hot-buffer residency not proven; Docker physical-core topology not verified
PostgreSQL 17.10; image sha256:270f9b60c903bac41b8587d5e912525cc87aa8381e3c52ca5bcc1206d7da4eda
pgs3 16 workers, 1 GiB shared buffers, 64 KiB direct threshold, default 4 MiB staged chunk size
Durability/checkpoints fsync/full-page-writes/synchronous-commit on; max_wal_size=4GB, checkpoint_timeout=30min; no workload checkpoint
MinIO RELEASE.2026-08-04T00-00-00Z, pinned one-server/one-volume image
Workload persistent HTTP/1.1 SigV4, deterministic distinct payloads, three warmups, zero retries

No normalization was applied. HTTP honors pgs3.chunk_size for staged PUT and snapshots it per request across SIGHUP. D027 removes the ordinary staged PutObject completion reread only after the restricted worker and PostgreSQL have bound an exact ordered canonical chunk manifest. Earlier one-request diagnostic samples measured 119.24 MiB/s at 2 MiB, 154.92 MiB/s at 4 MiB, and about 32 MiB/s at an 8 MiB allocation/storage cliff. The complete 4 MiB-default sweep now measures 167.894 MiB/s for the required 8 MiB PUT and is authoritative.

Complete object-size sweep

Milliseconds, operations/s, and MiB/s; nearest-rank percentiles.

Size System PUT p50 PUT p95 PUT p99 PUT ops/s PUT MiB/s GET p50 GET p95 GET p99 GET ops/s GET MiB/s Errors
4 KiB pgs3 5.779 15.652 27.553 1555.986 6.078 2.863 9.984 14.574 4065.195 15.880 0
4 KiB MinIO 9.130 14.565 16.995 1662.550 6.494 2.815 9.154 12.537 4137.809 16.163 0
16 KiB pgs3 4.664 9.894 17.896 2443.748 38.184 3.342 8.501 11.448 3989.933 62.343 0
16 KiB MinIO 9.007 14.019 15.997 1701.106 26.580 5.218 13.350 18.613 2577.770 40.278 0
65,535 B pgs3 4.694 15.049 23.559 1932.459 120.777 3.328 8.549 11.280 4055.658 253.475 0
65,535 B MinIO 10.669 15.227 17.362 1456.363 91.021 3.890 9.726 13.202 3530.574 220.658 0
64 KiB pgs3 2.731 9.421 13.103 3646.459 227.904 2.150 5.507 7.440 6154.713 384.670 0
64 KiB MinIO 9.038 12.999 15.556 1692.439 105.777 3.024 7.364 9.759 4575.654 285.978 0
65,537 B pgs3 8.159 15.365 23.382 1328.234 83.016 3.523 9.673 14.297 3631.641 226.981 0
65,537 B MinIO 10.106 15.430 19.618 1495.458 93.468 3.105 7.470 10.317 4467.902 279.248 0
256 KiB pgs3 8.562 25.839 41.856 1033.729 258.432 2.811 5.563 7.178 5029.700 1257.425 0
256 KiB MinIO 8.993 14.267 16.575 1661.785 415.446 2.450 5.725 6.739 5501.952 1375.488 0
1 MiB pgs3 31.891 89.721 96.305 284.384 284.384 5.321 23.999 25.755 1802.875 1802.875 0
1 MiB MinIO 13.847 22.077 25.940 1054.880 1054.880 4.552 9.043 11.373 2927.418 2927.418 0
4 MiB pgs3 42.862 76.687 90.170 64.593 258.372 8.774 15.064 15.291 369.104 1476.415 0
4 MiB MinIO 22.301 26.624 26.650 176.297 705.186 3.305 4.471 4.651 1093.349 4373.397 0
8 MiB pgs3 47.400 56.605 56.605 20.987 167.894 39.466 41.363 41.363 99.145 793.157 0
8 MiB MinIO 20.826 21.027 21.027 49.161 393.291 5.912 9.261 9.261 533.925 4271.401 0
16 MiB pgs3 90.509 102.809 102.809 10.647 170.348 51.463 73.798 73.798 31.641 506.251 0
16 MiB MinIO 35.795 38.154 38.154 27.601 441.615 9.964 11.217 11.217 191.992 3071.865 0
64 MiB pgs3 395.761 421.511 421.511 2.447 156.602 444.676 511.342 511.342 2.087 133.553 0
64 MiB MinIO 196.367 198.684 198.684 5.062 323.951 50.358 73.855 73.855 15.756 1008.384 0

Acceptance checks

Requirement Target Final result Status
Gate 12 small GET p50 <0.5 ms max 3.341708 ms FAIL
Gate 12 small GET rate >=30,000/s min 3989.933/s FAIL
Gate 13 small PUT rate >=5,000/s min 1555.986/s FAIL
Gate 14 8 MiB PUT >=150 MiB/s 167.894 MiB/s, p50 47.399667 ms, client concurrency 1 PASS
Gate 15 LIST / delimiter LIST <5 / <10 ms 2.358 / 9.487 ms PASS
Gate 16 fork 100k <1 s 1944.071 ms FAIL
Gate 17 complete comparison complete 11 sizes, both systems, zero errors PASS

Gate 14 used one persistent client connection against a 16-worker listener pool. Sequential keep-alive keeps that socket on one accepting worker, although the evidence does not record the worker PID alongside every sample.

The golden-image LIST plans touch 1,020/5,139 shared blocks and emit zero WAL. Fork correctness, refcounts, and independent mutations pass; fork emits 154,178,689 WAL bytes. The same workspace digest on pre-golden image sha256:3a83ef8c... measured fork/LIST/delimiter at 865.495/1.086/5.238 ms, all PASS. That iteration demonstrates that the materialized DML CTE and pgs3.blob fillfactor 80 optimization can meet the gate, but the golden-image 1944.071 ms result is authoritative and keeps gate 16 failed.

Findings

  • D026's boundary remains visible: 64 KiB direct PUT p50 is 2.731 ms versus 8.159 ms one byte above on the staged path.
  • D027 plus the 4 MiB default raises the required 8 MiB PUT from the earlier failed 29.436 MiB/s observation to 167.894 MiB/s on the final sweep. This passes gate 14 without changing durability settings.
  • fork_bucket feeds inserted blob IDs through one materialized DML CTE into a set-wise refcount update, and pgs3.blob uses fillfactor 80. Correctness passes, but final-image timing remains unstable and fails gate 16.
  • A CACHE 1024 experiment on object_version_id_seq was rejected after it broke externally observed version ordering in one selected Ceph case (194/195). The final schema retains PostgreSQL's default CACHE 1, and the golden Ceph run passes 195/195.
  • Repeat on a documented power-safe local NVMe reference host before production sizing; never normalize away these Docker Desktop results.