rmw_wait rounds its timeout to whole milliseconds: 1 kHz timers skip periods and short waits time out early #82

Generated by Generative AI

Claude Code (Claude Opus 5.5)

Operating System:

Linux 6.8.0-142-generic #142-Ubuntu SMP PREEMPT_DYNAMIC x86_64 GNU/Linux

ROS 2 distro:

jazzy

rmw_unix_socket_cpp version or commit hash:

devel @ 6d9793c

Client library (if applicable):

rclcpp

Executor (if applicable):

SingleThreadedExecutor

Does the same reproducer work with another RMW?

Yes — it works with another RMW, so the bug looks specific to rmw_unix_socket_cpp

QoS profile (if applicable):

No response

'ros2 doctor --report' output

No response

Steps to reproduce issue

1. Run a node with a 1 kHz timer (create_wall_timer(1ms)) that counts its callbacks for 30 s, with RMW_IMPLEMENTATION=rmw_unix_socket_cpp.
2. Or call rmw_wait directly on a guard condition that is never triggered, with a 0.3 ms and then a 1.7 ms timeout, and time each call.

Expected behavior

• About 30 000 timer callbacks in 30 s, as with Fast DDS, Cyclone DDS and Zenoh (29 998 to 30 001).
• rmw_wait returns RMW_RET_TIMEOUT no earlier than the timeout, and shortly after it.

Actual behavior

• 28 174 of 30 000 callbacks: about 6 % of the periods are skipped (29 622 with SCHED_FIFO).
• A 0.3 ms wait takes about 1.06 ms.
• A 1.7 ms wait returns RMW_RET_TIMEOUT after about 1.06 ms, before its deadline (20 of 20 tries).
• Knock-on effects: a 1 kHz request/reply loop between two processes loses 2–7 % of its replies, and the iRobot cli_srv 10 B benchmark completes only 35 310 of 40 000 calls in 20 s.

Additional information

Cause: src/rmw_wait.cpp converts the timeout to whole milliseconds for epoll_wait, in three places:

• the caller's timeout is truncated (nsec / 1000000), and anything under 1 ms becomes 1 ms ("At least 1ms");
• the deadline is computed from that rounded value, which is why a 1.7 ms wait ends after 1 ms;
• each time the wait re-blocks, the remaining time is rounded up to the next millisecond.

rclcpp passes the time until the next timer with nanosecond resolution, so a 1 kHz timer almost always asks for less than 1 ms and wakes up late.

PR #8 changed this code for infinite timeouts and kept the <1ms → 1ms rounding on purpose. This issue is about that rounding.

Proposed fix: keep the timeout in nanoseconds and block with epoll_pwait2() (Linux 5.11, glibc 2.35). Where it isn't available (older glibc, ENOSYS, or EPERM from a seccomp profile), fall back to epoll_wait() with the remaining time rounded up.