Performance Enhancement: Parallelized Unpacking of Velodyne Packets

[mebasoglu]

Feature Overview

This enhancement introduces a significant performance improvement by parallelizing the process of unpacking Velodyne packets. By leveraging parallel processing techniques, the repository's functionality will experience enhanced efficiency and reduced processing time during the unpacking stage. This improvement is particularly relevant for VLS-128 where large amounts of Velodyne packet data need to be processed quickly and efficiently.

Could you please provide any feedback, suggestions, and guidance on the proposed enhancement @drwnz @amc-nu ?

Here are the performance metrics obtained with profiling scripts:

	nebula-vls128-main	nebula-vls128-parallel
d_total AVG	24.8819	6.36068
n_out AVG	213395	213395
d_total STD	9.54071	2.84753
n_out STD	114.332	114.378
d_total AVG % rel to nebula-vls128-main	100	25.5634
n_out AVG % rel to nebula-vls128-main	100	99.9999
d_total STD % rel to nebula-vls128-main	100	29.8461
n_out STD % rel to nebula-vls128-main	100	100.041

[drwnz]

@mebasoglu looks interesting, are you able to create a PR for your implementation?

[mebasoglu]

Hello @drwnz , I created the PR: #62

[mojomex]

Hi @mebasoglu, we implemented the per-packet decoding proposed in

• Enhancement: allow dynamic scan message size to reduce latency #94
• Improve the overall performance #109

in

• refactor(velodyne)!: combine Velodyne ROS wrappers into a single node #148

and will thus close this issue in favor of that approach. The reason we favored the other approach is that packets arrive one-by one anyway, and instead of accumulating them to be able to decode in parallel, decoding them in a single thread as they arrive is even more memory efficient. it is also faster from a latency perspective, as upon arrival of the last packet, only one packet has to be decoded until the pointcloud can be published.

Nevertheless, thank you for your investigation and proposal.