Performance¶
Archived benchmark results. They were produced by
benchmarks/egernia-performance, a cluster benchmark harness this repository
used to carry, which measured the service, PostgreSQL as the data grows,
replica scaling and autoscaling behaviour separately — because they fail
separately.
That harness has been removed in favour of the timing guards in the
integration suite and the CPU microbenchmarks (see
Performance testing). The measurements below stand as
records; the make benchmark-* targets and egernia_bench commands the pages
reference no longer exist, and the harness itself is in the git history.
Each cluster run below is published in full: the graphs, per-measurement CSV, and provenance needed to know whether two runs are comparable at all. Focused capacity checks record smaller local experiments separately.
Focused capacity checks¶
- 20260831T141551Z-a042537-slow-readers-listings — two-slot slow-reader retention and 100k-row collection listings
Latest cluster run¶
20260825T180219Z-f8b21fc4-worker-sweep · CSV
Latest by family¶
| family | run | datasets |
|---|---|---|
db-scaling |
20260825T005436Z-b450b0a9-db-scaling | D1 D2 D3 D4 |
fixed-scaling |
20260824T014320Z-a5058118-fixed-scaling | D2 |
keda |
20260824T140134Z-bf7e4b24-keda | D1 |
profile |
20260825T163627Z-034d69ba-profile | D1 |
tap-compare-scaling |
20260912T225808Z-d43b538d-tap-compare-scaling (the three-server scaling comparison: egernia vs DaCHS vs argus at 8/16/24 CPUs — the paper's figure) · 20260905T073915Z-5f145831-tap-compare-scaling (egernia vs DaCHS, superseded) | D1 |
tap-compare |
20260911T065452Z-41bfc513-tap-compare (the final three-way comparison, Table A — one-process parity: egernia vs DaCHS vs argus) · 20260912T025434Z-41bfc513-tap-compare (Table B — equal CPU: the same three servers, egernia with eight workers) · 20260908T061813Z-f1cf2c25-tap-compare (egernia vs argus, equal CPU) · 20260906T145338Z-7aa14f6a-tap-compare (egernia vs argus) · 20260903T160103Z-8fec0e75-tap-compare (egernia vs DaCHS) | D1 |
worker-sweep |
20260825T180219Z-f8b21fc4-worker-sweep | D1 |
All runs¶
- 20260912T225808Z-d43b538d-tap-compare-scaling — the three-server resource-scaling comparison at 8 / 16 / 24 CPUs: egernia scales ×3.6, DaCHS and argus ×1.3
- 20260912T025434Z-41bfc513-tap-compare — the final three-way comparison, Table B: egernia (eight workers) vs the same GAVO DaCHS and CADC argus, unchanged
- 20260911T065452Z-41bfc513-tap-compare — the final three-way comparison, Table A: egernia (one worker) vs GAVO DaCHS vs CADC argus, all three on disjoint cpusets
- 20260908T061813Z-f1cf2c25-tap-compare — egernia (8 workers) vs CADC argus, equal-CPU parity
- 20260906T145338Z-7aa14f6a-tap-compare — egernia vs CADC argus, one-process parity
- 20260903T160103Z-8fec0e75-tap-compare
- 20260901T054529Z-3c8add85-tap-compare
- 20260825T180219Z-f8b21fc4-worker-sweep
- 20260825T163627Z-034d69ba-profile
- 20260825T005436Z-b450b0a9-db-scaling
- 20260824T140134Z-bf7e4b24-keda
- 20260824T102832Z-29507cbb-keda
- 20260824T074332Z-b4fa9d64-keda
- 20260824T014320Z-a5058118-fixed-scaling
- 20260823T231057Z-f0cd5bd7-db-scaling
- 20260823T184745Z-b16f52f5-db-scaling
Reading these numbers¶
- A figure without an interval is one measurement. Intervals across repetitions are 95% Student-t; percentile intervals within a run are percentile bootstrap.
LOAD_GENERATOR_BOUNDon any measurement means the client was the limit, and nothing else from that measurement describes the service.- A run marked invalid is published anyway, with the reason. The samples are the evidence of what went wrong.
- Two throughputs closer together than the run-to-run variability plot shows are the same throughput.