Repository navigation
bench: follow-ups from the 2026-10-22 corpus campaign (WebKit, 2 MiB pages, 2/8 vCPU memory and throughput, scale end-to-end test) #1173
Description
Follow-ups from the 2026-10-22/23 corpus campaign (#1172), each left out of that PR to keep its box time near one day.
Measurements not taken:
- WebKit corpus run (
make bench-chromium-corpus ENGINE=webkit ARMS=cdp,cdp-fast,noop,noop-fast CPU=4, about 40 min). Its golden refused itself on a harness bug fixed in report: measure what the Chromium render page called not measured (2026-10-22/23 corpus campaign) #1172 (873d12c3); the rerun was not done. - 2 MiB pages on the corpus: a minor, copy, minor A/B/A on a
HUGEPAGES=1golden, about 90 min. - Memory phase and open-loop throughput at 2 and 8 guest vCPUs, about 3 h per vCPU count.
- A third serial run at 2 and 8 vCPUs, and a third prefetch on/off pair, about 2.5 h together.
Harness:
- The scale path had four defects that only a real run found (REVIEW.md items 15 to 18 in report: measure what the Chromium render page called not measured (2026-10-22/23 corpus campaign) #1172): the request function ignored the cgroup guard and the fault probe, the probe raced fcvm's helpers, and the analyzer could not read the port mappings or the audit roles the harness writes. Each side had tests against its own fixtures. A test that drives
reqscale.pyend to end on a fake fcvm and feeds its directory toreqscale_analyze.pywould have caught all four. reqscale.py's rate rules (whole requests per ramp, whole corpus cycles per window) are checked only after the golden and the diag; see bench: corpus scale run keeps up at 1.4 req/s per restore path and misses at 1.87 with the host idle #1159.
Activity
Record the 2026-10-26 follow-ups: WebKit corpus, 2 MiB A/B/A, third s…
ejc3 commentedon Oct 26, 2026
Open, from the review of #1340: the diag now exempts a request the page cancelled before any network timing (ERR_ABORTED, no address, no timing) from the 'every traced request names an allowed address' rule. Reasoning for the exemption is in the #1340 thread (all corpus names resolve to the replay, so such a request can only have addressed it). Narrowing it needs the trace to record that the request was sent (cdpdrive.py: Network.requestWillBeSent / requestWillBeSentExtraInfo), which would let a cancelled request with a sent marker be held to the address rule again.
ejc3 commentedon Oct 26, 2026
Measured; record and numbers in #1347 (REVIEW.md section 'Follow-ups from the 2026-10-22 corpus campaign, run 2026-10-26'). Done: WebKit corpus (cdp 1,689 ms, 0 failures in 230), 2 MiB minor/copy/minor A/B/A (412/377/398 ms; 4 KiB comparison is a different golden), third serial runs at 2 and 8 vCPUs, memory per clone (screenshot) at 2 and 8 vCPUs (300 and 290 MiB cgroup, 4 vCPUs 340), third prefetch pair (+33%). Not done / open: HTML-extraction memory at 2 and 8 vCPUs (refused by the 2x cross-basis gate, unexplained), throughput ceilings at 2 and 8 vCPUs (runs above 1.4 per path are not analysis-passing; failures above ~4 per path are harness limits, see #1340), the cancelled-request diag exemption (open item above).