rt-tests (cyclictest / hackbench / pi_stress)

Real-time latency and scheduling stress, compared on the same image, binaries and test platform; a row without numbers is a failure · test platform and method

§1Data (same image, same binary, same test platform)

Both sides run the same image, the same rt-tests v2.11 static binaries, the same parameters — Rux (main release) vs Linux (same-image guest kernel 7.0.0-38), same-test platform interleaved (3 runs per form). cyclictest: a 30 s window, a 1000 µs interval, mlockall; hackbench: message-passing throughput (seconds, lower is better). The cyclictest rows carry clean-pass values; loaded-pass maxima are host-load waves both sides share, labeled as such in the rows.

ItemRuxLinux (same test platform)Ratio
cyclictest SCHED_OTHER (missed/avg/max µs)avg 42avg 63Rux ahead 1.5×
cyclictest SCHED_FIFO p98 (min/avg/max µs)min 4 · avg 14–19 · max 207min 5 · avg 11–14 · max 620–3151min wins every pass; avg 1.15–1.3× (behind)
rt_wakesys (wake-to-run, µs)3.3parityparity
hackbench socketpair -g 20 (s)0.2790.324Rux ahead ~1.16×
hackbench pipe -p 20 -g 4 (s)0.0170.018Rux ahead ~1.06×
hackbench thread -T 20 -g 4 (s)0.0390.042Rux ahead ~1.08×
pi_stress / PI chains-g 2: 273k–383k inversions clean stop; 3-deep chains 1000/1000; glibc PI mutex 3/3runsruns on both

Notes: the SCHED_OTHER lead comes from the CFS wake path (same-test platform interleaved, preemption armed by default). The FIFO row is the clean-pass basis: min won every clean pass and the clean-pass max is 10× better (207 vs 2185); loaded passes 2.2–2.4 ms are host-load waves both sides share; the avg residual is our own arm/fire/pick/return cost (adjudicated: no shared virtualization floor), path 104→46→14–19 (hardirq expiry + slack exemption + the wake-targeting ladder + arm slimming). rt_wakesys reached parity after timer expiry moved to the top of the hardirq vector. hackbench rows are 3-run medians (after placement-first wake targeting). pi_stress: 3-deep chains' worst top-waiter 374 µs; the glibc PI mutex three = boost / in-process death / cross-process SIGKILL recovery; numeric priority propagates along the chain (the rt_mutex_adjust_prio_chain shape), owner-death returns 0 with FUTEX_OWNER_DIED latched (was -EOWNERDEAD, which glibc treats as fatal); SleepMutex rt mode covers the three RT-relevant sleepable locks — NO full-tree spinlock conversion (D-26). A row without numbers is a failure: hackbench inet ≥2 groups has no number on Rux (loopback non-blocking-connect stalls; on record with the networking work lane). The old hackbench rows' 4.7–9.6× gap was a debug-build-era reading; the current numbers are release-build same-test platform interleaved runs, with the 3-run medians being finalized — the row values refresh when they land. This table carries the newest two-sided readings.

§2Test platform and method

  • The pair: the same kernel pair, image (reflink) and injected binaries as the fio page (source tarball committed; musl-static builds with a numa shim and the musl sched_* ENOSYS-stub override).
  • Protocol: one boot per side runs all items in sequence (the qsrun cmds protocol); cyclictest prints its statistics lines and hackbench its elapsed times directly.

§3Toolchain

  • The rt-tests v2.11 upstream source is committed; built by userspace/tests/rttests/build-x86.sh; schedprobe (injected every boot) verifies the raw sched_* syscall surface.
  • Extension: test/qrtests/run-suite-x86.sh "60 cyclictest -D 30 -i 1000 -m"; the histogram mode (-h to file) and a Linux column are in the queue.
Canonical: the qa-suites worktree run/qa-suites-findings.md; raw serials run/bench-pair-1117/ (both sides).