Repository navigation
perf: cut ROS2 ingest CPU ~2x, decline WS deflate; Lyrical CI; dependency bumps - #13
Merged
Merged
Conversation
Binary frames are already zstd-compressed. When a client offered the extension, IXWebSocket deflated every frame again with zlib on the publish thread: 7-11% of process CPU in a perf profile, 19.2% -> 12.9% of a core for one client on an 83-topic workload. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Profiling showed >50% of process CPU in rcl_wait: rmw_fastrtps re-attaches every subscription to the wait set on each wait cycle, and blocking spin() does one cycle per received message. - Subscription callbacks now drain their reader, so one cycle harvests a whole burst (the executor hands over only one message per cycle). - The ingest executor polls every ingest_poll_interval_ms (default 5, 0 = old blocking spin). - min_qos_depth default 1 -> 10 so a poll interval cannot overflow a shallow reader. - Publish/request/timeout timers moved to a dedicated callback group and executor thread: zstd no longer blocks ingest. 83 topics, ~700 msg/s, 1 client: 7.5% -> 3.4% of a core, identical message counts. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…st poll - heavy_frame_zstd_level (ROS2 param / --heavy-frame-zstd-level): zstd is ~60% of CPU on point-cloud workloads. Default stays 1 (no behaviour change); measured on 4 lidars, 52 MiB/s in: 1 = 19% of a core, 31 MB/s out; -5 = 15%, 39 MB/s; -100 = 10%, 51 MB/s. Wire-compatible. - Ingest and serializer copies no longer zero the buffer first. - ROS2 ingest loop blocks instead of polling while there are no subscriptions, so idle CPU matches the old blocking spin (0.45%). - Validation, docs and changelog for the new options. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
- Drain moved into a GenericSubscription subclass (handle_serialized_message) instead of a callback holding a weak_ptr to its own subscription: the pointer was assigned after the subscription became executable, racing with the ingest thread. - Receive time now comes from the RMW received_timestamp when available; with batched delivery "now" would cluster timestamps and distort per-client rate limiting. - MessageBuffer::set_latched() seeds the retained sample from the buffer: with ingest on its own thread the sole latched sample can arrive before the subscriber's bookkeeping, and later clients would get no replay. - Timer thread is joined on every exit path and reports exceptions instead of terminating; take errors during drain are caught. Builds and passes subscription tests on Humble and Jazzy. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
ament_target_dependencies was removed in newer distros, so the package did not configure on Lyrical. Modern targets work on Humble, Jazzy and Lyrical. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
No robostack-lyrical channel exists, so there is no pixi/conda/AppImage variant for Lyrical; colcon CI and the Debian release are covered. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
All frames keep zstd level 1 as before; the knob was a CPU-vs-bandwidth tradeoff not worth the configuration surface. FastDDS/RTI entry points are back to identical with main. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…LS link IXWebSocket 12.x is fixes only (server header parsing, fd double close, DNS crash, move instead of copy in send buffering). pixi.lock is left alone: moving the Humble env to 12.x re-solves the whole RoboStack stack. The fetched static IXWebSocket does not export its OpenSSL dependency, so the standalone FastDDS/RTI link failed with TLS on (also on 11.4.6). Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
facontidavide
changed the title
…exit On Lyrical, rcl_logging_spdlog registers a periodic flush thread on the global spdlog registry and rcl unloads the library on shutdown. After the suites' repeated init/shutdown the thread resumed in unmapped code at exit (all tests passed, process returned -11). The bridge binary is unaffected. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Fixes driven by real profiling (perf + pidstat against
ros2 bag playof real bags); analysis and numbers are in the companion docs PR (#14,docs/CPU_OPTIMIZATION.md§2.0 / §2.9).rcl_waitrebuilding the wait set once per received message. Subscriptions now drain their reader (GenericSubscriptionsubclass), the ingest executor polls everyingest_poll_interval_ms(default 5,0= old blocking spin, blocks while there are no subscriptions so idle is unchanged), and publish/request/timeout timers run on their own executor thread so zstd never blocks ingest.min_qos_depthdefault 1 → 10. Receive timestamps now come from the RMWreceived_timestamp.websockets; the PlotJuggler plugin does not offer it).MessageBuffer::set_latched()seeds from an already-buffered sample (race exposed by threaded ingest; failing test first).target_link_librariesinstead ofament_target_dependencies(removed in Lyrical), newROS: LyricalCI workflow, Lyrical.debin the release matrix. No pixi/conda/AppImage for Lyrical: there is norobostack-lyricalchannel..debbuild; fixes only, several server-side), Fast DDS 3.4.0 → 3.4.3, CLI11 2.6.0 → 2.6.2. Fixed a pre-existing link failure of the standalone backends when IXWebSocket is fetched with TLS.pixi.lockintentionally untouched: moving the Humble env to IXWebSocket 12 re-solves the whole RoboStack stack — separate chore.Measurements
i7-13700H, bridge pinned to P-cores,
performancegovernor, 1 client, losslessness verified by per-topic message counts decoded client-side.Dead end worth knowing: polling without draining silently loses data (executor takes one message per subscription per wait cycle; 481 Hz
/tffell to 154 Hz).Test plan
rmw_cyclonedds_cpp, not installed)c57a313.debjob runs only on tags / manual dispatch — not exercised by this PR🤖 Generated with Claude Code