# hp physical-NIC revalidation — 2026-09-30 This is a hardware test of the C++ transport paths, not an A/B verdict on a latency optimisation or a replacement for the multi-flavour load matrix. ## Provenance and method - Source for the rebuilt FIX-session binaries: `96f6d487cce53df44385051bdad28164805952c9`. GCC 14.2.1, `-O3 -flto=auto -march=native -mtune=native`, profile-free (`tuned_build=0`). The installed C++ FIX binaries were replaced from those builds, and their hashes matched the build outputs. - Host: `hp`, Xeon Gold 6254, SMT off, Linux `5.14.0-687.44.1.el9_8.x86_64`; isolated server/client cores 8/9. VMA's maintenance core is reserved separately as 13 by the hp machine file. - Solarflare: Onload `dde226d7 2026-08-07`; Mellanox: libvma `9.8.80-4.el9`. Both namespace rigs reported link and passed their physical port preflight. - Each of the six arms per card ran serially for three repeats, each with 20,000 warm-up and 200,000 measured round trips. `ping_pong` and the full FIX-session `bench_roundtrip` both ran kernel, preload and native-stack arms. Every repeat emitted exactly 200,000 samples. The figures below are the median of the three reported per-repeat percentiles, in microseconds; they are not percentiles recomputed from pooled raw samples. - The matrix launcher was repaired before this run. It now fails on missing binaries, client errors, missing or short sample rows, and too-small physical TX deltas. Its previous post-arm `verify` call sent its *own* ping, so that check could pass even if an Onload arm had stayed off the wire. The new check brackets **each** arm with physical-port counter reads, excluding the verifier's own ping packets. | Card | Workload | Transport | p50 | p99 | p99.9 | Physical TX delta during arm | | --- | --- | --- | ---: | ---: | ---: | ---: | | Solarflare | wire floor | kernel | 11.512 | 12.390 | 13.700 | 1,320,013 | | Solarflare | wire floor | Onload | 3.504 | 4.055 | 4.199 | 1,320,013 | | Solarflare | wire floor | TCPDirect | 3.107 | 3.178 | 3.260 | 660,010 | | Solarflare | FIX session | kernel | 12.790 | 13.646 | 14.461 | 660,019 | | Solarflare | FIX session | Onload | 4.852 | 5.219 | 5.378 | 660,016 | | Solarflare | FIX session | TCPDirect | 4.237 | 4.420 | 4.599 | 660,013 | | Mellanox | wire floor | kernel | 11.087 | 13.500 | 14.689 | 1,320,013 | | Mellanox | wire floor | VMA | 3.379 | 3.564 | 4.901 | 1,320,013 | | Mellanox | wire floor | SocketXtreme | 3.339 | 3.922 | 5.306 | 1,320,015 | | Mellanox | FIX session | kernel | 12.794 | 14.277 | 16.415 | 660,019 | | Mellanox | FIX session | VMA | 5.470 | 6.024 | 6.212 | 1,320,018 | | Mellanox | FIX session | SocketXtreme | 5.117 | 5.726 | 7.087 | 1,320,020 | The counter deltas include the benchmark warm-up and measured traffic, but not the counter-check pings. These are same-host, separate network namespaces joined by a DAC. They establish live physical traffic, not cross-host latency. ## Other NIC checks - Solarflare `test_fix_efvi_live` opened a live VI on `ens5f0`, received five multicast UDP payloads from the opposite namespace, and returned a non-zero hardware RX timestamp (`997693667112223`). Its sender reported 10,000/10,000 datagrams sent. - Mellanox `test_fix_ibv_live` opened a raw-packet QP on `ens4f0np0`, received five matching payloads, and returned a non-zero hardware RX timestamp (`155891328070331`). Its sender also reported 10,000/10,000 sent. - The current-source `test_fix_zf_lifecycle_fake` passed, including the case that skips packet timestamp lookup after an idle reactor pass. The current-source `nictime-recv-test` passed both live receive paths with its 64-byte control buffer. The installed ef_vi/ibverbs test executables were built in August; their source has not changed since then. The full FIX-session binaries above were freshly built for this run. The installed `bench-run` dispatcher differed from current source only in comments during measurement and was refreshed afterwards. ## Historical commands and recovered artifacts The commands below describe the 2026-09-30 run, which predates the new session-worktree policy. Future runs must be launched from their own `~/code/wt//libhft` checkout, with build products under `~/build//libhft` where supported and retained evidence under `~/archive/libhft/`; the shared `~/code/libhft` checkout is pull-only. ```sh make -C engine/cpp bench_roundtrip zf_session vma_session sudo -n /usr/local/sbin/libhft-bench-install bench_roundtrip sudo -n /usr/local/sbin/libhft-bench-install bench_roundtrip_zf sudo -n /usr/local/sbin/libhft-bench-install bench_roundtrip_vma_sx REPS=3 ITERS=200000 WARMUP=20000 RIG_NIC=sfc private/ops/scripts/run-wire-matrix.sh REPS=3 ITERS=200000 WARMUP=20000 RIG_NIC=mlx5 private/ops/scripts/run-wire-matrix.sh ``` The actual runs set `OUT` to `/tmp`. Those files had disappeared by the time the workspace was migrated. The two per-repeat CSVs below were reconstructed from the exact lines captured in the session output and verified against the SHA-256 values recorded at run time. They are byte-identical to the original CSVs, but are recovered artifacts, not files moved from `/tmp`. | Artifact | SHA-256 | | --- | --- | | `~/archive/libhft/nic-20260930/libhft-wire-sfc-final-20260930.csv` | `b9c9ddbcc95a990f153e6c5a6b53986862e9018be96b24c604efc2e140353e93` | | `~/archive/libhft/nic-20260930/libhft-wire-mlx5-final-20260930.csv` | `96b8e9fd3520b4944fe62e9aa60cac3737b58abb85e5ac12716efe3c7065f2e5` | The full command logs and the sender/receiver logs were **not** recovered. The original full-log hashes were `89ab04cad37cce436e999959d7240675cbb81f9e41658c6864b5b2bda736289b` (Solarflare) and `a45d939123885d0245b1f2454f7493b2a361a3fc125856c55ae2a8eeca5884a1` (Mellanox); these identify missing files, not currently available artifacts. No latency candidate was changed or promoted from this test. Any optimisation claim needs a pinned before/after build pair, interleaved repeats, raw samples and a host-interference audit. Java, .NET and Rust NIC load cells were not part of this C++ transport revalidation.