Skip to content

bench: publish verified three-node store-versus-peer E11 results - #83

Merged
gspark-etri merged 9 commits into
mainfrom
exp/82-multinode-peers
Oct 31, 2026
Merged

gspark-etri merged 9 commits into
mainfrom
exp/82-multinode-peers

Conversation

@gspark-etri

@gspark-etri gspark-etri commented Oct 31, 2026 •

Copy link
Copy Markdown
Contributor

Closes #82.

Measure the existing file-only B4/S256→B4/S512 transition on three eligible 1GbE Kubernetes nodes without accelerator resources. Four cache/arm conditions each have five simultaneous three-node cohorts (20 main, 60 node samples); four smokes and both setup attempts are retained. All target full-file hashes pass, no main retries/exclusions, and all owned CPU Jobs/Pods and the dedicated namespace are removed by UID.

With one seed in both arms, origin pack payload drops 395,822→0 B. Metadata remains 69,693,294 B, so all successful origin response bodies drop only 0.564740%. Empty peer snapshots provide no peer payload. Observed cohort medians store→peer: empty 3.102→3.240 s; one-seed 3.296→3.014 s. Report includes IQR and per-node times; no general latency speedup claim. Origin colocation, static snapshots, uncontrolled OS/load, instrumented proxy and application-byte scope are explicit.

Raw: 128-file package, tag ccgrid-e11-20261031-raw-v1; measured harness 7870cef872e201a6a4dca87ed1bb998da9c3109f, runtime 05c97605b07882f9ceaaeb90b94878f8923f81cf. Report, summary and PDF/SVG/PNG figure included.

Validation: seven accounting/validity regressions; offline verifier reconstructs all cohort counters from 26,928 original request rows, checks controller copies, cache/variant and source identities, actual node assignment and UID cleanup. Independent evidence review and clean Git archive reproduction pass; corruption is rejected. CI now reproduces the committed summary in addition to Go vet/race/build and real S3 conformance/CLI suites. Model payloads are absent from Git; offline checks compare worker-recorded full hashes with frozen inventories. No runtime behavior changes; NUMBERS.md untouched.

@gspark-etri gspark-etri changed the title experiment: measure three-node file delivery with and without peers bench: publish verified three-node store-versus-peer E11 results Oct 31, 2026
@gspark-etri
gspark-etri marked this pull request as ready for review October 31, 2026 02:39

Copy link
Copy Markdown
Contributor Author

Final verification at head ca0049c88014101a866ea4b2d39333fcb179eacc:

  • Independent whole-change/evidence review: no remaining Critical/Important/Minor findings; per-node hash-validation overlap wording corrected.
  • Clean Git archive of exact final head: seven regressions pass, all128 raw files rehash, full original-request reconstruction passes, reproduced summary equals committed JSON; checksum corruption rejected. Independent review also rejected request-byte, output-SHA, cache-placement and actual-node corruption.
  • CI 41161011264: all three checks green and actual job logs read. Go vet/race/build, seven regressions, raw summary reproduction/diff and both real S3 conformance+CLI round trips passed. CI merge checkout includes exactly this head over main05c9760.
  • Raw tag-v1 preserves measured-source ancestry; later final-head commit only clarifies report wording. No runtime/raw changes.
  • All4 final CPU Jobs/Pods succeeded then UID-deleted; final namespace absent and cdx Job/Pod inventory empty; no accelerator requested. NUMBERS.md untouched.

Proceeding with the authorized GitHub-flow squash merge.

@gspark-etri
gspark-etri merged commit d56fe3e into main Oct 31, 2026
3 checks passed
@gspark-etri
gspark-etri deleted the exp/82-multinode-peers branch October 31, 2026 02:42
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

E11: concurrent multi-node variant delivery, store-only versus peers

1 participant