Skip to content

benchmarking: add substrate-row SLO-knee sweep (gvisor + microvm) - #885

Open
Alex Bulankou (AlexBulankou) wants to merge 1 commit into
agent-substrate:mainfrom
AlexBulankou:a4s1/benchmarking-slo-ladder-tests
Open

benchmarking: add substrate-row SLO-knee sweep (gvisor + microvm)#885
Alex Bulankou (AlexBulankou) wants to merge 1 commit into
agent-substrate:mainfrom
AlexBulankou:a4s1/benchmarking-slo-ladder-tests

Conversation

@AlexBulankou

Copy link
Copy Markdown
Collaborator

Adds automation/tests-slo-ladder.yaml: a dedicated on-demand concurrency ladder (5 rungs × sandboxClass: gvisor|microvm, 10 entries) for SLO-knee measurement — deliberately separate from the standing tests.yaml so the sweep is invoked explicitly, not hourly. The file header documents the two per-class result globs for the SLO-knee post-processor.

Config-only; no cluster is touched by merging this.

tests-slo-ladder.yaml is a dedicated, on-demand tests file (NOT the standing
hourly tests.yaml) for a substrate-row SLO-knee measurement run. Point the
orchestrator at it via --tests.

It sweeps the warm-path AteAPIUser at a concurrency ladder (1/5/10/25/50
users) once per sandbox class (gvisor + microvm) -- 10 rungs total. Together
the rungs form the throughput/latency sweep that
benchmarking/analysis/slo_knee.py reduces to a knee (max sustained RPS whose
chosen latency percentile stays under the SLO ceiling).

Kept out of the hourly tests.yaml on purpose: each rung triggers a full
substrate deploy + teardown, so this is a deliberate on-demand sweep.

Analysis convention baked into the header: stats partition by test NAME
(runs/<name>/...), and both classes share tag+metric, so the two ladders are
analysed as separate name-prefix globs (slo_gvisor_* vs slo_microvm_*).
Passing --tests-yaml back at this file feeds slo_knee the per-rung user count
(verified: users_map_from_tests_yaml parses all 10 rungs).

[a4s1] reference patch for #6595 benchmarking spike.

Signed-off-by: Alex Bulankou <alexbu@google.com>
@a4-a4s1
a4-a4s1 Bot force-pushed the a4s1/benchmarking-slo-ladder-tests branch from 01e187f to f602013 Compare August 12, 2026 09:16
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