Implement lock free ring buffer with throttling [HUMBLE/IRON]

[Samahu]

Related Issues & PRs

• Closes Missing lidar chunk os1-128-u  #240
• Closes Parts of point cloud missing #272
• Related Atomic refactor for lock-free ringbuffer [Foxy] #302
• ROS1[NOETIC] Implement lock free ring buffer with throttling [NOETIC] #319
• ROS2[FOXY] Implement lock free ring buffer with throttling [FOXY] #320

Problem Analysis

• When adding multiple subscribers or when having a bad network or low performing system, LidarScan composition and processing could eventually take more time, this leads to a build up in the packets buffer which will eventually be exhausted and the current mechanism of the thread_safe_ring_buffer::write_overwrite made things only worse. The solution can be in one of two things either ensure that the LidarScan composition + processing takes less time than the time needed to add new packets, or decouple packet addition from LidarScan consumption, if the consumption thread can't keep up then throttle.

Summary of Changes

• Implement lock free ring buffer with unit tests
• Decouple packet reception from packet processing
• Throttle the ring buffer as we reach 70% capacity
• Drop packets only when the ring buffer has been completely exhausted

TODO(s):

• Replace the LockFreeRingBuffer implementation with ThreadSafeRingBuffer within os_sensor_nodelet.
  • Alternatively, forward the packets directly to LidarScan without a staging buffers.

Validation

• Run the driver as usual with RIVZ, add multiple subscribers to /ouster/points
• Verify the problem with missing columns is eliminated or reduced.
• Observe that the point cloud publish frequency isn't affected significantly

[TillBeemelmanns]

We tested both, this branch and this attempt #302 to solve the problem with missing parts of the point cloud. We set the recv buffer size to the following values

sudo sysctl -w net.core.rmem_default=26214400
sudo sysctl -w net.core.rmem_max=26214400

on the host machine and also set proc_mask:="PCL". However in all cases we could observe partial missing point clouds as soon as we add a second subscriber node to the driver. That applied to resolutions 1024x20, 2048x10, 4096x5, for the other resolutions 512x10, 512x20, 1024x10 we could observed a significant delay in the point cloud without visible package loss.

For all resolutions, the point cloud looks fine if only RVIZ is used for visualization (single subscriber).
The host is a high performance workstation with 64 cores and 128 GB RAM.

[marioney]

We also tested this PR with an OS-1-128 Rev 7 and it fixed the missing points issue for us, running ROS Humble in Ubuntu 22.04.

We tested both 1024x10 and 1024x20 modes and the issue was not present in any of those.

[Samahu]

@TillBeemelmanns @marioney thank you both for the feedback.

@TillBeemelmanns sorry that didn't help with your situation .. I still have one TODO item that I predict it to help with the delay and might improve things overall. I should have this PR completed and merged by next week.

[mkiael]

Tested on OS0-128 Rev C running ROS Iron on Ubuntu 22.04. Solves issue with missing data.

Great work!

[Samahu]

Thanks @mkiael for the feedback. I will proceed with the PR as is since it resolves the issue for most of the driver users and will address the rest of suggested improvements in a separate PR.