Repository navigation
perf-callgrind: JVM vs clojure-rs per-namespace sweep with Bencher #1
Description
Mirror of snippet-repo T1211 (hhh-pa/clojure-code-snippet#37): a deterministic callgrind sweep over every test namespace, run under both JVM Clojure and clojure-rs, reporting callgrind_instructions + callgrind_allocated_bytes per benchmark to Bencher (same benchmark names, distinct testbeds).
Proven already:
- Single-ns driver works under clojure-rs:
alter *load-path* -> require -> clojure.test/run-testsprints the standard summary (Needs: avoidSystem/exit, unemulated — encode pass/fail in output/exception). - Recipe (valgrind flags, JVM -Xint+EpsilonGC, LD_PRELOAD alloc shim, BMF shape) ports from the snippet repo's
scripts/ci/go/internal/measure.
Design deltas vs T1211 (spec to follow on the branch):
- Benchmarks = 1351 test namespaces (not script x cases); driver
driver.clj <ns>. - Correctness gate = SUMMARY line (
:test/:pass/:fail/:errorequal on both legs, zero fail/error) instead of full-stdout diff (namespaces print nondeterministically). - New Bencher project +
BENCHER_API_KEYsecret (admin step); newperf-callgrindworkflow (shard x runtime matrix) +report-callgrindmerge;testunchanged.
Session: gentle-menkar.
Activity
Issue hhh-pa#1: algoperf tool, driver, workflow + 4 test-isolation fixes
Issue hhh-pa#1: workflow review fixes (gate name, JDK pin, action ver…
Issue hhh-pa#1: locale pin, count-diff exceptions, review fixes
m-hu commentedon Nov 3, 2026
PR #2 (draft, branch
m-hu:perf/1-callgrind) is up: spec atdocs/1/specification.md, full implementation, both 1351-leg sweeps green locally. Draft until theBENCHER_API_KEYadmin step lands.Session: gentle-menkar.
AI-Tool: Muse Code 1.4.14-R7309.1 (Meta Muse Spark)
Issue hhh-pa#1: sweep skip-list, relaxed shim check, OOM hint
Issue hhh-pa#1: skip problem-056 (7.4g EpsilonGC need)
perf-callgrind: JVM vs clojure-rs per-namespace sweep with Bencher (#3)
m-hu commentedon Nov 3, 2026
Session: gentle-menkar. Landing record: delivered by #3 (squash ab26ab9 onto main) — perf-callgrind JVM-vs-rs per-namespace sweep with Bencher tracking, green on CI + all 16 perf shards with uploads proven. Draft #2 superseded; follow-up flake in #4.