Speed up C++ logging for many individual log calls #4287
Assignees
Labels
Milestone
Comments
Wumpf
added
🚀 performance
4 tasks
Wumpf
added a commit
that referenced
this issue
Nov 22, 2023
…ing a component type registry (#4296) ### What * Fixes #4287 * Follow-up to #4273 As expected, not doing the C++ datatype -> C FFI schema -> Rust datatype roundtrip for each log call helps perf quite a bit, especially when we do a lot of smaller log calls. The registry a single RwLock protected Vec (we never deregister) which is exposed via a single c entry point. On the C++ side we use the local `static` variable mechanism for threadsafe lazy registration (slight codegen adjustment). Indicator components had some special handling before and were refactored to fit in this system - in the process I made their arrow array shared across all instantiations, further cutting down on per-log work. --- Benchmark results: * large point cloud: `0.15s` -> `0.14s` * many points: `7.52s` -> `4.52s` * large images: `0.57s` -> `0.51s` Old values from previous PR. New values are median over three runs, single executable run (this makes more and more of a difference with all these registries!), timings without prepare step, same M1 macbook. A quick look over the profiler for running `log_benchmark points3d_many_individual` in isolation tells us that of the actual benchmark running time we spend.. * 35% of the the time in `rr_recording_stream_log` (of which in turn 20%, so 7% overall, is still arrow FFI translation of the array!!) * 30% in the various `to_data_cell` methods * 10% in exporting arrow arrays to C FFI * 6% in setting the time * the rest in various allocations along the way (taken via `Instruments` on my Mac) <img width="969" alt="image" src="https://github.com/rerun-io/rerun/assets/1220815/5632589f-52b1-4e92-b7a0-1482e69528ad"> --- ### Checklist * [x] I have read and agree to [Contributor Guide](https://github.com/rerun-io/rerun/blob/main/CONTRIBUTING.md) and the [Code of Conduct](https://github.com/rerun-io/rerun/blob/main/CODE_OF_CONDUCT.md) * [x] I've included a screenshot or gif (if applicable) * [x] I have tested [demo.rerun.io](https://demo.rerun.io/pr/4296) (if applicable) * [x] The PR title and labels are set such as to maximize their usefulness for the next release's CHANGELOG - [PR Build Summary](https://build.rerun.io/pr/4296) - [Docs preview](https://rerun.io/preview/8bf1ee59d9a2bc5e192c1c8169c98dd40b621100/docs) <!--DOCS-PREVIEW--> - [Examples preview](https://rerun.io/preview/8bf1ee59d9a2bc5e192c1c8169c98dd40b621100/examples) <!--EXAMPLES-PREVIEW--> - [Recent benchmark results](https://build.rerun.io/graphs/crates.html) - [Wasm size tracking](https://build.rerun.io/graphs/sizes.html)
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
The recent performance improvement
got the C++ sdk a lot faster. But compared to Rust we're still behind for individual log calls (like time series scalars!).
An obvious candidate to improve is not building & sending the schema every time: Right now on every log call we convert the schema to C FFI and then create a Rust/arrow2 representation from it. Add a simple lazy schema registry/handle system for this!
Should do a little bit more profiling though to get an idea where the perf goes. E.g. there's likely many many other needless allocs on the way.
The text was updated successfully, but these errors were encountered: