Hi Ben, Thanks for the patch. We'll build 46956 on our 26.06 base and run it in the lab.
A note on what we can measure: on our production routers, three days after a restart, show clock error is in the nanoseconds, so the drift builds slowly there. Our lab VM drifts faster (−34 ms after a few days), so it will be the main testbed. We'll report error over time with and without the patch. One question: 46956 removes the clib_max() ratchet in clib_time_verify_frequency() , but the worker offset clamp in vlib_worker_thread_barrier_check() stays. Baldur flagged both. Is the slewing expected to keep each thread close enough to main that the barrier clamp never engages, or does that need changing too? On LACP: understood, the time source is the real fix. Since 26.06 and 26.10 will ship without it, I'll post the lacp_time_now() change to gerrit for Steven to decide on, as defence in depth. Happy to abandon it once 46956 is in. Steven (Steven Luong, LACP maintainer): I'll add you as reviewer on that change. In short, lacp-input arms the periodic timer with the worker's vlib_time_now(), but lacp_periodic() checks it against main's clock, so any worker/main skew delays or blocks LACPDU transmission (Baldur's #27196 above). The change gives LACP one CLOCK_MONOTONIC timebase on all threads. Your view on whether you want it in the plugin would be appreciated. Regards, Pieter Meyer
-=-=-=-=-=-=-=-=-=-=-=- Links: You receive all messages sent to this group. View/Reply Online (#27226): https://lists.fd.io/g/vpp-dev/message/27226 Mute This Topic: https://lists.fd.io/mt/121360743/21656 Group Owner: [email protected] Unsubscribe: https://lists.fd.io/g/vpp-dev/leave/14379924/21656/631435203/xyzzy [[email protected]] -=-=-=-=-=-=-=-=-=-=-=-
