Skip to content

ci(testnet): contract load window after funding, block rates from the first block - #75

Merged
EnzinoBB merged 5 commits into
masterfrom
ci/load-fixes
Oct 23, 2026
Merged

EnzinoBB merged 5 commits into
masterfrom
ci/load-fixes

Conversation

@EnzinoBB

@EnzinoBB EnzinoBB commented Oct 23, 2026 •

Copy link
Copy Markdown
Collaborator

Harness fixes from Fable's analysis of the slowdowns in the scale runs (#72). None of them were node issues.

  • Load window (H1). The window was set before funding N writer accounts, which takes one block each. At 9 nodes that left about 100 s of the 120 s for the load, and at 51 nodes only about 30 s, which is the only reason the achieved rates fell with N. The window now starts after the funding.
  • Writers in step (H2, tried and reverted). Spreading the writers over their period made every call land in its own block, where it queues behind the previous call's new state, because one contract runs one execution cycle at a time (about 6 rounds). Write p50 went to 12.6 s at 9 nodes and 37 s at 51 nodes, against 3.0 and 4.8 s with the writers in step. Calls in the same block share one cycle, so the writers start together again.
  • Block rates (H3). The sampler computes per-interval block rates from the first block on. The start, with the executors' version check, is not a slowdown.

Validated by scale run 39876040185 (9 and 51 nodes, PASS):

🤖 Generated with Claude Code

EnzinoBB and others added 4 commits October 22, 2026 22:37
…pread over their period, block rates from the first block

From the slowdown analysis of the scale runs:
- load(): the window was set before funding N writer accounts (a block each), leaving ~100 s of
  120 at 9 nodes and ~30 s at 51: the achieved rates fell with N for that reason only
- the writers started together with the same period and stayed in lockstep: every write landed in
  one block with the others (51 calls executed one by one), which made write latency look like it
  grew from 3.6 s to 4.8 s at 51 nodes; they now start spread over the period, with jitter
- sampler: per-interval block rates from the first block on (the start, with the executors'
  version check, is not a slowdown)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…k run in one cycle)

Spreading the writers made every call land in its own block and queue behind the previous call's
new state: write p50 12.6 s at 9 nodes and 37 s at 51 (51 refused), against 3.0 and 4.8 s in step.
One contract runs one execution cycle at a time (about 6 rounds); calls of the same block share it.
The load window after funding and the block rates from the first block stay.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@EnzinoBB EnzinoBB changed the title ci(testnet): contract load window after funding, staggered writers, block rates from the first block ci(testnet): contract load window after funding, block rates from the first block Oct 23, 2026
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@EnzinoBB
EnzinoBB merged commit 7499055 into master Oct 23, 2026
2 checks passed
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.

1 participant