lmbench 3.0-a9

Latest state: the authoritative same-test platform KVM table (12 latency + 8 bandwidth rows, each row the newest measured value on current main) · Canonical: docs/test/lmbench-report.md

§1lmbench data (same-test platform KVM, 21 cases)

Source: the "Standing (2026-11-16, kernel main=017bf556b)" section of lmbench-report.md — the single authoritative table; the data below is that section. Method: both sides in the same test platform (q35/KVM/-cpu host/smp2/2G, host CCD0 pin + performance governor + warmer thread); Rux boots with nohz_full=1, one case per boot with the retry rule; the Linux column runs the OHOS 6.6.101 kernel in the same guest test platform with the same static binaries, 2 reps, medians. The switch/wakeup family is sensitive to the Linux-side measurement window — ratios are only meaningful within one session.

Basis: each row carries the latest measured value on current main (the 2026-11-16 full round; the null/read/write rows re-measured 11-17). The Rux column follows each round's own build class; the Linux column = the host 7.0.0 kernel booted in the same guest test platform running the same static binaries.

Ahead / parity (19 cases)

CaseRuxLinux (same test platform)Ratio
lat_pipe4.40 µs25.06–26.27 µsahead 5.84×
lat_ctx -s64 22.51 µs12.84–12.97 µsahead 5.14×
lat_ctx -s64 81.92 µs12.06–12.08 µsahead 6.29×
bw_pipe30,464 MB/s8,241 MB/sahead 3.7×
lat_sem0.503 µs4.47–4.55 µsahead 8.97×
bw_unix48,352 MB/s23,022 MB/sahead 2.1×
lat_unix2.26 µs3.24 µsahead 1.4×
lat_syscall fstat0.084–0.094 µs0.119–0.120 µsahead 1.3×
bw_mem 4m bzero71,489 MB/s58,245 MB/sahead 1.23×
bw_mem 1m bcopy84,520 MB/s74,252 MB/sahead 1.14×
lat_mem_rd 1MB2.20–2.25 ns/load2.375 ns/loadahead 0.95×
lat_mem_rd 256MB6.49–6.65 ns/load6.691 ns/loadparity-plus (0.97×)
lat_usleep 10001030.8 µs1052.8 µsparity (0.98×)
bw_file_rd 1m io_only33,408–35,545 MB/s23,045 MB/s (ceiling)ahead 1.45–1.54×
lat_pagefault0.0253 µs0.072 µs (ceiling)ahead 2.8× (vs ceiling)
bw_mmap_rd 1m mmap_only79,954 MB/s77,100 MB/s (ceiling)ahead 1.04×
lat_syscall null0.042 µs0.046 µsahead 0.92×
lat_proc fork36.6–40.3 µs33.9 µsparity band 1.08–1.19×
lat_proc exec39.4–40.6 µs37.6 µsparity band 1.05–1.08×

Behind (7 cases)

Note: lat_fs is the largest true gap on the board (file create/delete hot path); the former load-coupled stall is fixed (three stacked roots: the sleep exemption wiping live interrupt-nesting counts, softirq counts migrating across tasks, tmpfs page allocation under a spinlock), and the table carries the latest post-fix reading; lat_connect's fabricated ETIMEDOUT is fixed; bw_tcp is functionally complete (all six grid cells pass, typical 5 s) with an occasional bimodal slow mode in the same family as lat_fs (open); bw_mmap_rd and pagefault recovered after the image-runner fix. Note: lat_connect's fabricated ETIMEDOUT is fixed (established-then-closed must return success) — 2000/2000 reproducer + net LTP spot 12/12; bw_tcp narrowed: a bare 2MB bulk transfer verifies complete and fast, the stall sits in the benchmark's own inter-block protocol exchange (open).

CaseRuxLinux (same test platform)Ratio
lat_syscall open0.676 µs0.394–0.397 µsbehind 1.71×
lat_syscall read0.095 µs0.060–0.062 µsbehind 1.55×
lat_syscall stat0.27 µs0.239 µsbehind 1.13× (after the stat-2x lane)
lat_syscall write0.078 µs0.0597–0.0600 µsbehind 1.30×
lat_tcp 127.0.0.110.35 µs4.89 µs (ceiling)behind 2.12×
lat_udp 127.0.0.19.00 µs4.147 µs (ceiling)behind 2.17×
lat_fs -s 1k20.1k/17.3k files/s445k/455k (ceiling)behind 22–26×
lat_connect / bw_tcpfixed66.5 µs (ceiling)connect() bug fixed

Notes: the bw_file_rd and lat_pagefault rows carry values recovered after the image-runner fix (the earlier lows were not trustworthy); the lat_syscall null flip comes from the fdget borrow + getppid O(1) + PT_REGS + fdreap silence + the NOHZ stack. Mechanism note: the wins concentrate in scheduler/wakeup and IPC (cross-vCPU wakeup IPIs are expensive under virtualization; the single global run queue produces no cross-core reschedule storm) and memory bandwidth; syscall-latency rows run the guest's own entry (no vmexit), so the gaps there are real kernel differences. Rows without honest numbers: bw_mmap_rd and the net-server family (lat_tcp/udp/connect, bw_tcp, lat_fs) — not listed; lat_mmap 4.525 µs (1.12x vs ceiling, near-parity band).

Correctness bugs the campaign found and fixed (the performance gaps were hiding them): a stale-rsp0-at-free UAF (the fork bistability root cause); a call_rcu zero-grace callback firing; a wake-monitor staleness inversion; the fdreap per-tick 10 ms close stall; the mainline munmap final-TLB-flush loss (5b4a4bc7).
Why the large leads: design trade-offs, not missing safety features.Code-level argument, item by item:
(1) Source of the wake/switch family gap (sem 8.97x / ctx-8 6.29x / ctx-2 5.14x / pipe 5.84x) — the reference's per-CPU runqueues under KVM make every ping-pong roundtrip pay cross-core resched IPIs (vmexit + halt-exit, 2-5 µs each; the IPI send side alone measured 935-1139 ns). Rux's single global run queue (register D-1) plus the 120 µs polling idle window turns µs-scale handoffs into same-core, ZERO-IPI transfers (sched/sched.rs idle window, closegap lane measurements). This is queue topology winning under virtualization's cost structure.
(2) The safety machinery is fully present on exactly these paths: the polling window's lost-wake impossibility is by protocol (flag-clear + SeqCst fence + recheck Dekker pairing, sched.rs idle loop); every enqueue carries wakewatch forensic stamps and schedstat wait accounting; all GRQ mutations run under the INV-LOCK-* lock discipline; RCU quiescent-state reporting runs at BOTH syscall entry and exit (sync/rcu.rs).
(3) User-memory access safety matches the reference shape: every uaccess (scalars included) goes through exception-table fixups (arch/x86_64/uaccess.rs) — bad pointers return EFAULT, never crash. x86 SMAP is currently not armed (honestly recorded, sec-r1), but its performance contribution to these cases is ~2 two-cycle instructions per copy (~1 ns) — 0.005% of the pipe gap's 20 µs scale — not the source of the lead; the semantics are preserved by the active exception-table path.
(4) fork/exec copy-on-write and teardown safety present: fork's mmfork phase carries the full PTE write-protection (COW, per the forkprof decomposition); page-table freeing goes through an RCU grace period; fd lifetimes go through the retire queue (fs/fdreap.rs).
(5) The bug taxonomy corroborates: every memory-corruption-class historical bug lived in audited unsafe code, zero in safe code (docs/architecture/safety-and-performance.md §2). The full lead list with measured evidence: docs/architecture/rust-native-register.md (D-numbers and same-test platform A/B figures).

§3Reproducing

# the full suite (one boot per case + the retry rule)
cd userspace/lmbench && ./extract-lmbench.sh && ./build-lmbench.sh && ./build-rootfs.sh work/lmbench-x86.img && ./run-rux.sh
# the quick path (10 headline cases, ~3 minutes, marked INDICATIVE)
./quick.sh [bzImage]
Permanent image store: /home/william/rux-images/testimages/lmbench-x86.img (reflink-copied, booted snapshot=on — the store copy is never mutated). Canonical: docs/test/lmbench-report.md.